apolinario Claude Code commited on
Commit
f453775
·
1 Parent(s): 9b31755

Refresh all metrics, add reflex's launch post, raise the recency credit

Browse files

Metrics: re-read every item against X, GitHub (via the authenticated gh CLI,
5,000/hr instead of 60) and the Hub, 43 of 46 entries updated. The snapshot
moves to 2026-09-19 in the footer, README and BASE_DATE.

The refresh reshaped the top of the list: Laya went 163 -> 1,410 GitHub stars
and 108 -> 577 Hub likes (raw 1,679 -> 11,666), which lifted it above LocalJev
and dropped LocalJev from #3 to #5 at the old credit of 5000. The credit is
raised to 10000, which puts LocalJev back at #2 with a 926-point cushion over
#4 rather than the 51-point one a credit of 9000 would give -- a margin thin
enough that the next refresh would flip it back. Top 5 is now Laya, LocalJev,
45-second TL;DR, Qwen-2.5-1B-RLCD, openjev.com.

Note the engagement gate now saturates for 20 of 46 items, so for anything
with real traction the credit acts as a flat recency tier and only age
separates items. That is the intended behaviour here -- it is what keeps a
0-engagement entry off the list -- but it does mean the credit is no longer
"small", which was the original constraint.

reflex (Kshetrajna Raghavan, @kshetrajna ) was already listed; this adds its
launch post and WebGPU demo, a real description of what shipped since (a
/v1/systemone server, temperature scaling to 0.039 ECE against Jev's 0.031,
and an open LoRA recipe), its author name, and fresh metrics (330 likes,
138k views, 77 stars).

Also corrected typesafe-ai-benchmark to its actual tweet date (2026-09-17),
found by comparing every entry against its tweet's created_at.

README.md and og.html (plus the re-rendered og.png, now showing LocalJev at
#2) are updated to match.

Co-Authored-By: Claude Code <noreply@anthropic.com>

Files changed (5) hide show
  1. README.md +2 -2
  2. index.html +50 -49
  3. og.html +50 -49
  4. og.png +2 -2
  5. refresh.js +9 -6
README.md CHANGED
@@ -22,7 +22,7 @@ This static Space tracks the artifacts in one combined grid, color-coded by kind
22
  - **Prior art**: "this already exists" claims
23
  - **Explainers**: architecture speculation, explainers, benchmarks and roundups
24
 
25
- Cards sort by a **trending** score: ♥ likes on X + 5 × GitHub stars + 8 × Hub likes + views ÷ 500, plus a **recency credit** (up to 5000, halving daily) so a fresh release is not buried under older, louder ones. The credit is scaled by the item's own engagement, so a brand-new entry nobody has reacted to yet cannot climb on its date alone. Cards added within a day of the metrics snapshot get a green **new** badge. Use the category and "has" chips to filter, or switch the sort.
26
 
27
  …plus a section on what is still **not** in the open: TypeSafe's weights, the RLCD algorithm, and any open model matching Jev's calibration claims.
28
 
@@ -36,7 +36,7 @@ Hugging Face models and Spaces are first-class artifacts here, not only GitHub r
36
 
37
  All data lives in one JS array (`ITEMS`) near the bottom of `index.html`. Open a PR on this Space to add or correct an entry. Please include the announcement post on X if there is one, and the Hub / GitHub link.
38
 
39
- Engagement metrics are a snapshot (2026-09-18) from public post metadata and will drift.
40
 
41
  ## Credits
42
 
 
22
  - **Prior art**: "this already exists" claims
23
  - **Explainers**: architecture speculation, explainers, benchmarks and roundups
24
 
25
+ Cards sort by a **trending** score: ♥ likes on X + 5 × GitHub stars + 8 × Hub likes + views ÷ 500, plus a **recency credit** (up to 10000, halving daily) so a fresh release is not buried under older, louder ones. The credit is scaled by the item's own engagement, so a brand-new entry nobody has reacted to yet cannot climb on its date alone. Cards added within a day of the metrics snapshot get a green **new** badge. Use the category and "has" chips to filter, or switch the sort.
26
 
27
  …plus a section on what is still **not** in the open: TypeSafe's weights, the RLCD algorithm, and any open model matching Jev's calibration claims.
28
 
 
36
 
37
  All data lives in one JS array (`ITEMS`) near the bottom of `index.html`. Open a PR on this Space to add or correct an entry. Please include the announcement post on X if there is one, and the Hub / GitHub link.
38
 
39
+ Engagement metrics are a snapshot (2026-09-19) from public post metadata and will drift.
40
 
41
  ## Credits
42
 
index.html CHANGED
@@ -315,7 +315,7 @@
315
  <div>
316
  <p><b>Missing something?</b> Open a PR against this Space's <code class="mono">index.html</code> (the data is one JS array at the bottom of the file) or ping <a href="https://huggingface.co/multimodalart" target="_blank" rel="noopener">@multimodalart</a>. Hugging Face models and Spaces count as first-class artifacts here, not just GitHub repos.</p>
317
  <p><b>Trending</b> = ♥ likes on X + 5 × GitHub stars + 8 × Hub likes + views ÷ 500. All categories share one grid, color-coded; use the category chips to filter. Promised-but-unreleased items always sort last.</p>
318
- <p>Engagement numbers are a one-time snapshot from public X post metadata on 2026-09-17 and will drift. GitHub stars and Hub likes were read the same day. Descriptions paraphrase each project's own README and launch post; corrections welcome.</p>
319
  <p>Visual identity and outlined Huggies from <a href="https://huggingface.co/spaces/Chunte/HFBA" target="_blank" rel="noopener">HF Huggiverse (Chunte/HFBA)</a>. Unofficial, community-maintained, not affiliated with TypeSafe AI.</p>
320
  </div>
321
  <img src="huggies/rocket.png" alt="">
@@ -348,117 +348,118 @@ const ITEMS = [
348
  desc: 'Batches every JSON key at once and reads category probabilities from stock Qwen 2.5 1B logits, ~5× faster than generating JSON.',
349
  quote: 'They were building in stealth for 2 years, I was building in stealth for 2 hours… No new training required.',
350
  links: { tweet: 'https://x.com/harshagundal/status/2100044305536889015', hf: 'https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD', space: 'https://huggingface.co/spaces/drinkmoonshine/parallel-constrained-decoding' },
351
- m: { likes: 5507, rts: 318, views: 1145212, hfLikes: 343 }, date: '2026-09-16', tags: 'qwen json mac m4 on-device forks epsilon3 shreyansh26 drinkmoonshine' },
352
  { cat: 'decode', title: 'openjev-sglang: Jev-compatible public API', by: { name: 'Eric Zhang', handle: 'ekzhang1' },
353
  desc: 'A drop-in /v1/systemone endpoint over Qwen3.6-35B-A3B. SGLang\'s radix cache keeps the shared prefill hot, so 64 parallel tasks finish in under a second. Prefill-only, no generation.',
354
  quote: 'It runs a comparable open model (Qwen3.6-35B-A3B), and just uses SGLang radix cache to preserve the prefill reuse — 64 tasks in <1s.',
355
  links: { tweet: 'https://x.com/ekzhang1/status/2100651678110515383', gh: 'https://github.com/ekzhang/openjev-sglang' },
356
- m: { likes: 763, rts: 42, views: 91918, stars: 115 }, date: '2026-09-17', tags: 'sglang qwen3.6 moe api systemone radix cache' },
357
  { cat: 'decode', title: 'openjev + openjev.com', by: { name: 'Theo Lee', handle: 'theoleecj' },
358
  desc: 'Reads option logits on Qwen3.5-4B instead of generating text: 21 decisions in 1.02s vs 5.33s for JSON. CUDA plus a WebGPU browser demo.',
359
  quote: 'On the Jev waitlist? You can use not Jev (but close enough) right here.',
360
  links: { tweet: 'https://x.com/theoleecj/status/2100243609060856264', gh: 'https://github.com/TheoLeeCJ/openjev', demo: 'https://openjev.com' },
361
- m: { likes: 10, rts: 0, views: 3043, stars: 1427 }, date: '2026-09-16', tags: 'webgpu browser rtx 3090 qwen3.5 logits mit' },
362
  { cat: 'decode', title: 'PocketJev: on-device iPhone visual decisions', by: { name: 'aqty', handle: 'aqty' },
363
  desc: 'Qwen3-VL running locally on an iPhone via MLX. Camera frame + question + 3 options → direct option logits in ~1s, no photos saved.',
364
  quote: 'iPhone上でQwen 2Bをローカル実行して、カメラ画像+質問+3択から文章生成なしで直接判定。',
365
  links: { tweet: 'https://x.com/aqty/status/2100549517171175807', gh: 'https://github.com/NullPo-jp/PocketJev' },
366
- m: { likes: 6, rts: 4, views: 351, stars: 0 }, date: '2026-09-17', tags: 'ios iphone mlx vision qwen3-vl camera japan' },
367
  { cat: 'decode', title: 'jevmlx: Jev-style decisions for any MLX model', by: { name: 'Beni', handle: 'beni_il_' },
368
  desc: 'Typed schema fields broadcast across the KV cache, one batch row per field, all scored in a single forward pass on any MLX model.',
369
  quote: 'If you can run Qwen on your machine, it can behave very similarly to Jev.',
370
  links: { tweet: 'https://x.com/beni_il_/status/2100617387116568956', gh: 'https://github.com/bnsd55/jevmlx' },
371
- m: { likes: 7, rts: 0, views: 1154, stars: 19 }, date: '2026-09-17', tags: 'mlx apple silicon kv cache hebrew' },
372
  { cat: 'decode', title: 'jev-on-a-laptop: the honest benchmark', by: { name: 'Richard Becker', handle: 'RichardBcker1' },
373
  desc: 'Qwen 2.5 1.5B/7B and Qwen3 8B on a laptop vs Jev\'s published cases: 7B hits 73.8% agreement to Jev\'s 86.6%, 0.4–2s per decision. Finds confidence does not reliably flag errors.',
374
  quote: 'All code, benchmarks, model-size comparisons and the full write-up (unofficial, no affiliation with TypeSafe).',
375
  links: { tweet: 'https://x.com/RichardBcker1/status/2100195276107530393', gh: 'https://github.com/rorshopping/jev-on-a-laptop', space: 'https://huggingface.co/spaces/rorshopping/parallel-constrained-decisions' },
376
- m: { likes: 1, rts: 0, views: 62, stars: 14 }, date: '2026-09-16', tags: 'benchmark study agreement fraud apple silicon m5 writeup explainer' },
377
  { cat: 'decode', title: 'jevfire: one context, many decisions on CUDA', by: { name: 'Enrique Alcázar', handle: 'enriquealcazarg' },
378
  desc: 'Parallel decisions for CUDA LLMs behind a vLLM-style API, with game-agent examples. Claims 10×+ over regular inference, runs in the browser, accepts images.',
379
  quote: 'Here you have an open-source approach to the same problem… uses the same principles as JEV, is type-safe, runs in your browser, and accepts images.',
380
  links: { tweet: 'https://x.com/enriquealcazarg/status/2100267330022375858', gh: 'https://github.com/kikoncuo/jevfire', site: 'https://kikoncuo.github.io/jevfire/' },
381
- m: { likes: 1, rts: 0, views: 246, stars: 4 }, date: '2026-09-16', tags: 'cuda vllm browser images games primeagen' },
382
  { cat: 'decode', title: 'LFM2.5-RLCD (350M & 2.6B): PCD on Liquid models', by: { name: 'notnotsamuel · monotykamary', handle: null },
383
  desc: 'Parallel constrained decoding on unchanged Liquid LFM2.5 weights, branching attention and convolution state. Inference-only: 8–63× faster, field accuracy still ~60%.',
384
  links: { hf: 'https://huggingface.co/notnotsamuel/LFM2.5-350M-RLCD', hf2: 'https://huggingface.co/monotykamary/LFM2.5-2.6B-RLCD' },
385
- m: { hfLikes: 7 }, date: '2026-09-16', tags: 'liquid lfm hub weights renamed inference only pcd' },
386
  { cat: 'decode', title: 'parallelConstraintDecoding (Java + Python)', by: { name: 'Stephan Janssen', handle: 'Stephan007' },
387
  desc: 'A Java port of the decode trick alongside a Python reference.',
388
  quote: 'Here\'s a java version of Jev 🤩',
389
  links: { tweet: 'https://x.com/Stephan007/status/2100673219967557906', gh: 'https://github.com/stephanj/parallelConstraintDecoding' },
390
- m: { likes: 3, rts: 0, views: 99, stars: 6 }, date: '2026-09-17', tags: 'java jvm port' },
391
  { cat: 'decode', title: 'system-one (snellingio): "System One Lite"', by: { name: 'snellingio', handle: null },
392
  desc: 'Choice / Score / Noul server over MLX Qwen3, mirroring TypeSafe\'s three primitives. No launch post found.',
393
  links: { gh: 'https://github.com/snellingio/system-one' },
394
- m: { stars: 28 }, date: '2026-09-16', tags: 'mlx qwen3 choice score noul server' },
395
  { cat: 'decode', title: 'open-jev (daseinlabs): Gemma 3 4B on MLX', by: { name: 'Dasein Labs', handle: null },
396
  desc: 'Prefix-cache option scoring on a local Gemma 3 4B via MLX, a Doom demo, and a /v1/systemone shim. Only seen in roundups.',
397
  links: { gh: 'https://github.com/daseinlabs/open-jev' },
398
- m: { stars: 25 }, date: '2026-09-17', tags: 'gemma mlx doom systemone shim' },
399
  { cat: 'decode', title: 'openvons (open-Jev, Japanese)', by: { name: 'genai-craft', handle: null },
400
  desc: 'Decision layer answering finite-option questions with probabilities, from text, images or Japanese voice commands.',
401
  links: { gh: 'https://github.com/genai-craft/openvons' },
402
- m: { stars: 7 }, date: '2026-09-17', tags: 'japanese voice image multimodal' },
403
  { cat: 'decode', title: 'openjev (zhihz): bilingual local decisions', by: { name: 'zhihz', handle: null },
404
  desc: 'Local bilingual probability decisions from context, questions and candidate answers.',
405
  links: { gh: 'https://github.com/zhihz/openjev' },
406
- m: { stars: 5 }, date: '2026-09-16', tags: 'chinese bilingual local' },
407
  { cat: 'decode', title: 'system-one (sgoedecke): batched single-token choice', by: { name: 'Sean Goedecke', handle: null },
408
  desc: 'Batched single-token choice inference for open models, wire-compatible with TypeSafe.',
409
  links: { gh: 'https://github.com/sgoedecke/system-one' },
410
- m: { stars: 2 }, date: '2026-09-17', tags: 'compatible batched single token' },
411
- { cat: 'decode', title: 'reflex: stock Qwen3.5 / Qwen3-VL behind /v1/systemone', by: { name: 'kshetrajna12', handle: null },
412
- desc: 'Runnable server for the "architecture unmasked" reconstruction: choice / noul / score over untouched Qwen3.5 and Qwen3-VL, shared state with isolated questions, optional image state. Still updating.',
413
- links: { gh: 'https://github.com/kshetrajna12/reflex' },
414
- m: { stars: 48 }, date: '2026-09-17', tags: 'systemone choice score noul midtraining qwen3-vl image shared state server' },
 
415
  { cat: 'decode', title: 'LitJev: Jev-schema decision layer for the Qwen family', by: { name: 'Zhengxu Yu', handle: 'yuzxfred' },
416
  desc: 'Shared prefill, forked KV/recurrent state, one cached branch per question, option logits read from lm_head on any Qwen checkpoint (default Qwen3.8-27B), behind the exact Jev /v1/systemone schema. Playground, MMLU-Pro eval, optional temperature calibration.',
417
  quote: 'Qwen3.8-27B,10 道 MMLU-Pro 一次完成,0.3 s。比让模型只生成答案快 2.9x。非官方,基于公开信息。',
418
  links: { tweet: 'https://x.com/yuzxfred/status/2100652136878981337', gh: 'https://github.com/zhengxuyu/litjev', blog: 'https://x.com/yuzxfred/status/2100896621173047367' },
419
- m: { likes: 2, rts: 0, views: 125, stars: 4 }, date: '2026-09-17', tags: 'qwen qwen3.8 27b systemone schema compatible calibration mmlu-pro doom chess vision transformers chinese explainer writeup limitations' },
420
 
421
  /* ---------------- trained ---------------- */
422
  { cat: 'decode', title: 'snapjudge: TypeSafe-compatible decisions on a Mac, scored against Jev', by: { name: 'Michael Gross', handle: 'michagross' },
423
  desc: 'Choice / Score / Noul behind /v1/systemone on MLX Qwen3.5-family models, probabilities read straight from the logits. On TypeSafe\'s 20 public cases (373 decisions) Qwen3.8-27B reaches 81.2% agreement vs Jev\'s 87.7%: level on short texts, behind on invoices.',
424
  links: { tweet: 'https://x.com/michagross/status/2101062157290856604', gh: 'https://github.com/Micha0827/snapjudge' },
425
- m: { stars: 2 }, date: '2026-09-18', tags: 'mlx apple silicon qwen3.6 qwen3.8 systemone api benchmark agreement german runner game' },
426
  { cat: 'trained', title: 'Bespoke Nimble: open data, open model, open recipe', by: { name: 'Mahesh Sathiamoorthy', handle: 'madiator' },
427
  desc: 'The first fully open recipe: LoRA on Qwen3.5-9B with the data curation published too. Contrastive curation flips one fact to make negatives, so training needs no probabilities. 90.1% agreement with Jev vs 66.4% for base Qwen.',
428
  quote: 'Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.',
429
  links: { tweet: 'https://x.com/madiator/status/2100990591215783946', gh: 'https://github.com/bespokelabsai/nimble', hf: 'https://huggingface.co/bespokelabs/Bespoke-Nimble-9B' },
430
- m: { likes: 163, rts: 16, views: 8393, stars: 13, hfLikes: 6 }, date: '2026-09-18', tags: 'bespokelabs nimble qwen3.5 9b lora contrastive data curation open recipe synthetic distillation free parallel scorer 100ms h100 macbook' },
431
  { cat: 'trained', title: 'jevlike: train your own one-pass option scorer', by: { name: 'Vinny LaRouge', handle: 'vinnylarouge' },
432
  desc: 'A tiny byte encoder (or frozen HF encoder) plus an option head maps context + N options to one probability each. Doom, chess and Wikispeedia demos; parent of several forks.',
433
  quote: 'I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes.',
434
  links: { tweet: 'https://x.com/vinnylarouge/status/2100170846346097083', gh: 'https://github.com/vinnylarouge/jevlike', demo: 'https://x.com/vinnylarouge/status/2100281651930513460' },
435
- m: { likes: 2006, rts: 168, views: 161595, stars: 818 }, date: '2026-09-16', tags: 'train scorer encoder doom chess wikispeedia rl' },
436
  { cat: 'trained', title: 'openjev (AlexWortega): Qwen3.5 + MLP head, real weights', by: { name: 'Alex Wortega', handle: 'justALEXWORTEGA' },
437
  desc: 'The only trained "openjev" checkpoint on the Hub: Qwen3.5-4B as an NLI cross-encoder, argmax entailment is the move. Plays Doom zero-shot from text and pixels; 35B-A3B heads included.',
438
  quote: 'Typesafe: pnewed 💨 Jev: liberated 🫡 I trained an MLP on top of qwen 4b and it works literally like JEV',
439
  links: { tweet: 'https://x.com/justALEXWORTEGA/status/2100341039986798930', hf: 'https://huggingface.co/AlexWortega/openjev', gh: 'https://github.com/mikesmullin/openjev' },
440
- m: { likes: 770, rts: 46, views: 101737, hfLikes: 113, stars: 1 }, date: '2026-09-16', tags: 'nli cross-encoder mlp qwen3.5 doom vision weights hub' },
441
  { cat: 'trained', title: 'decider-2b: full fine-tune of Qwen3.5-2B-Base', by: { name: 'Mapika', handle: null },
442
  desc: 'Full fine-tune of Qwen3.5-2B-Base on 942k examples from 64 decision datasets, plus a vision variant. Weights on the Hub.',
443
  links: { gh: 'https://github.com/Mapika/decider', hf: 'https://huggingface.co/Mapika/decider-2b', hf2: 'https://huggingface.co/Mapika/decider-2b-vision' },
444
- m: { stars: 86, hfLikes: 38 }, date: '2026-09-16', tags: 'fine-tune qwen3.5 2b calibrated vision gh200 hub weights' },
445
  { cat: 'trained', title: 'rlcd-modernbert-151m ("OpenJev Verdict")', by: { name: 'heman10x', handle: null },
446
  desc: 'Post-trained ModernBERT (151M) scoring 25 candidate slots, abstention included, in under 35ms. WebGPU playground, Apache-2.0.',
447
  links: { hf: 'https://huggingface.co/heman10x/rlcd-modernbert-151m', gh: 'https://github.com/Heman10x-NGU/Verdict-open-jev' },
448
- m: { hfLikes: 1, stars: 2 }, date: '2026-09-17', tags: 'modernbert encoder webgpu abstention gliclass hub weights' },
449
  { cat: 'trained', title: 'open-jev (pngwn): Qwen3.5-4B decision scorer + demo Space', by: { name: 'pngwn', handle: 'evilpingwin' },
450
  desc: 'Qwen3.5-4B LoRA fine-tune that scores each candidate answer for a typed question, with temperature scaling so confidence tracks accuracy. Weights (v1, v2b) and a demo Space on the Hub.',
451
  quote: 'Since everyone is open-jevving. I set ML Intern (on hugging chat) to work. After a few experiments and a little guidance and we did ok.',
452
  links: { tweet: 'https://x.com/evilpingwin/status/2100721265107751121', space: 'https://huggingface.co/spaces/pngwn/open-jev', hf: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer-v2b', hf2: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer' },
453
- m: { likes: 144, rts: 24, views: 8907, hfLikes: 4 }, date: '2026-09-17', tags: 'qwen3.5 scoring head temperature scaling calibration ml intern huggingchat hub weights space demo' },
454
  { cat: 'trained', title: 'jevbetter: a stronger jevlike', by: { name: 'olanotolu', handle: null },
455
  desc: 'jevlike with a fancier head: hashed n-gram encoder, rival-aware attention, temperature scaling. Benchmarked against jevlike.',
456
  links: { gh: 'https://github.com/olanotolu/jevbetter' },
457
- m: { stars: 10 }, date: '2026-09-16', tags: 'n-gram attention temperature scaling calibration' },
458
  { cat: 'trained', title: 'system-one-open: Gemma 4 E2B / Gemma 3 270M + head', by: { name: 'mithalouni', handle: null },
459
  desc: 'Typed calibrated decisions in one pass on Gemma 4 E2B and Gemma 3 270M, trained on Modal.',
460
  links: { gh: 'https://github.com/mithalouni/system-one-open' },
461
- m: { stars: 4 }, date: '2026-09-17', tags: 'gemma modal scoring head' },
462
  { cat: 'trained', title: 'qwen-rlcd: Qwen3.5-0.8B Choice/Score/Noul', by: { name: 'shamazharikh', handle: null },
463
  desc: 'Choice / Score / Noul on Qwen3.5-0.8B. Inference works; training marked on hold.',
464
  links: { gh: 'https://github.com/shamazharikh/qwen-rlcd' },
@@ -466,11 +467,11 @@ const ITEMS = [
466
  { cat: 'trained', title: 'system-one-gemma: Gemma 3 270M scoring head', by: { name: 'akash-kamat', handle: null },
467
  desc: 'Gemma 3 270M with a scoring head; single forward pass, no text generation.',
468
  links: { gh: 'https://github.com/akash-kamat/system-one-gemma' },
469
- m: { stars: 0 }, date: '2026-09-17', tags: 'gemma 270m tiny' },
470
  { cat: 'trained', title: 'Laya: 421M RLCD decision model + demo Space', by: { name: 'Convai Innovations', handle: null },
471
  desc: 'ModernBERT-large fine-tuned plus a decision head trained from scratch (421M total) with RLCD against proper scoring rules: choice/score/noul with calibrated probabilities, all questions in one pass. Weights and a 7-tab Gradio demo on the Hub.',
472
  links: { gh: 'https://github.com/NandhaKishorM/laya', hf: 'https://huggingface.co/convaiinnovations/laya', space: 'https://huggingface.co/spaces/convaiinnovations/laya-demo', blog: 'https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me', site: 'https://www.reddit.com/r/LocalLLaMA/comments/1wjieap/made_the_horizontal_opensource_model_for_jev_with/' },
473
- m: { stars: 163, hfLikes: 108 }, date: '2026-09-18', tags: 'modernbert rlcd calibrated choice score noul support triage guardrails rag moderation routing hub weights space demo github pypi reddit devto' },
474
  { cat: 'trained', title: 'System One Mini: 69M DistilBERT, five fixed decisions', by: { name: 'David Hatley', handle: null },
475
  desc: '69M DistilBERT with five fixed typed heads over synthetic software-diagnosis summaries. Its card says: not a reproduction of Jev or RLCD.',
476
  links: { hf: 'https://huggingface.co/DavidHatley/system-one-mini' },
@@ -478,86 +479,86 @@ const ITEMS = [
478
  { cat: 'trained', title: 'NanoJev: 0.6B Qwen3 with a real decision head', by: { name: 'TianyuCodings', handle: null },
479
  desc: 'A 0.6B Qwen3 backbone with an actual trained head emitting Choice / Boolean / Score across 2–255 candidates in one forward pass. Full training pipeline, maze and snake runs, weights and data on the Hub. The one Chinese and Korean roundups are pushing hardest.',
480
  links: { gh: 'https://github.com/TianyuCodings/NanoJev', hf: 'https://huggingface.co/C-Tianyu/NanoJev' },
481
- m: { stars: 235, hfLikes: 7 }, date: '2026-09-17', tags: 'nanojev 0.6b qwen3 decision head choice boolean score maze snake training pipeline weights hub china korea' },
482
  { cat: 'trained', title: 'mini-jev: a preregistered measurement of the trick', by: { name: 'r-ms', handle: null },
483
  desc: 'Not another clone: a preregistered study of letter-logit versus grammar-JSON decoding on frozen Qwen3-4B and CLINC150. An actual measurement, no weights trained.',
484
  links: { gh: 'https://github.com/r-ms/mini-jev' },
485
- m: { stars: 13 }, date: '2026-09-17', tags: 'preregistered study letter logits grammar json clinc150 qwen3-4b measurement' },
486
 
487
  { cat: 'trained', promised: true, title: 'Archer Hume\'s open-weight Jev (promised)', by: { name: 'Archer Hume', handle: '4rcherhume' },
488
  desc: 'Announced alongside the architecture teardown: an open-weight version trained on the inferred design. Training in progress, nothing released yet.',
489
  quote: 'I decided I\'d release an open-weight version. BUT training takes time, so while we all wait I decided I\'d drop the sauce.',
490
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
491
- m: { likes: 1982, rts: 182, views: 297794 }, date: '2026-09-17', tags: 'promised coming soon open-weight training archer' },
492
 
493
  /* ---------------- prior ---------------- */
494
  { cat: 'prior', title: 'GLiNER 2.5: "Weights are on Hugging Face."', by: { name: 'George (Fastino / GLiNER)', handle: 'george_onx' },
495
  desc: 'Points the launch thread at the GLiNER 2.5 collection: a pretrained zero-shot encoder-classifier already doing single-pass classification over caller-supplied labels.',
496
  quote: 'We\'ve been reading the comments. Weights are on Hugging Face.',
497
  links: { tweet: 'https://x.com/george_onx/status/2100293114808119379', hf: 'https://huggingface.co/collections/fastino/gliner25-models' },
498
- m: { likes: 1565, rts: 91, views: 219615 }, date: '2026-09-16', tags: 'gliner fastino zero-shot encoder classifier collection' },
499
  { cat: 'prior', title: 'DeepMost sales-conversion RL model + papers', by: { name: 'Nandakishor M', handle: 'Nandakishorm1' },
500
  desc: 'Claims the idea shipped in 2025 papers (arXiv 2503.23303, 2510.01237) with a Hub model and dataset. PPO over embeddings; loosely related.',
501
  quote: 'Also the second work published in September 2025 was exactly the same one jev proposed now.',
502
  links: { tweet: 'https://x.com/Nandakishorm1/status/2100451360873611666', hf: 'https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning', paper: 'https://arxiv.org/abs/2510.01237' },
503
- m: { likes: 0, rts: 0, views: 22, hfLikes: 86 }, date: '2026-09-17', tags: 'ppo embeddings arxiv prior art sales' },
504
 
505
  /* ---------------- arch ---------------- */
506
  { cat: 'explain', title: 'Jev\'s Architecture Unmasked', by: { name: 'Archer Hume', handle: '4rcherhume' },
507
  desc: 'Argues Jev is a causal transformer with shared-state encoding, per-question branches and direct probability readouts, likely on a sparse MoE backbone, inferred from latency probing.',
508
  quote: 'All the grifters are completely wrong about Jev\'s architecture so I decided I\'d release an open-weight version. BUT training takes time…',
509
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
510
- m: { likes: 1982, rts: 182, views: 297794 }, date: '2026-09-17', tags: 'moe teardown speculation open-weight promise blog' },
511
 
512
  /* ---------------- diffusion ---------------- */
513
  { cat: 'diffusion', title: 'DiffusionGemma-as-Jev + vLLM structured-generation patch', by: { name: 'Matt Mastracci', handle: 'mmastrac' },
514
  desc: 'Adds a Jev-like structured mode for DiffusionGemma to vLLM (PR #57250) and runs live evals: not slower than Jev\'s API and roughly tied on quality.',
515
  quote: 'Is Jev faster than DiffusionGemma? No ❌ Is Jev smarter than DiffusionGemma? No ❌ (they\'re roughly tied!)',
516
  links: { tweet: 'https://x.com/mmastrac/status/2100626193943052784', pr: 'https://github.com/vllm-project/vllm/pull/57250' },
517
- m: { likes: 575, rts: 36, views: 36393 }, date: '2026-09-17', tags: 'diffusion gemma vllm dgx spark evals' },
518
  { cat: 'diffusion', title: 'razorback16/openjev + api.codiv.ai host', by: { name: 'razorback16', handle: null },
519
  desc: 'The same vLLM PR #57250 route, packaged as a Jev-wire /v1/systemone service: a DiffusionGemma 26B-A4B canvas with only the label slots left as noise, one read-only denoise step reads out the distribution. Public host at api.codiv.ai.',
520
  links: { gh: 'https://github.com/razorback16/openjev', site: 'https://api.codiv.ai/v1/systemone' },
521
- m: { stars: 11 }, date: '2026-09-17', tags: 'diffusiongemma 26b a4b nvfp4 vllm 57250 hosted api codiv systemone denoise canvas' },
522
  { cat: 'diffusion', title: 'open-jev (JoshuaSP): typed JSON with DiffusionGemma', by: { name: 'Joshua SP', handle: null },
523
  desc: 'Typed JSON inference with DiffusionGemma, benchmarked against Every and Jev.',
524
  links: { gh: 'https://github.com/JoshuaSP/open-jev' },
525
- m: { stars: 14 }, date: '2026-09-16', tags: 'diffusion gemma benchmark every' },
526
  { cat: 'diffusion', title: 'LocalJev (GitHub Next): oMLX Jev API + model bake-off', by: { name: 'GitHub Next', handle: 'GitHubNext' },
527
  desc: 'Official GitHub Next bridge, since not everyone there has Jev access: a Bun server putting oMLX DiffusionGemma behind /v1/systemone. Prompts for JSON probabilities instead of reading logits: wire-compatible, not equivalent. Ships a 1,200-request bake-off.',
528
  quote: 'Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man\'s Jev on top of omlx for local use.',
529
  links: { tweet: 'https://x.com/GitHubNext/status/2101193436816920798', gh: 'https://github.com/githubnext/localjev' },
530
- m: { likes: 1084, rts: 148, views: 142175, stars: 328 }, date: '2026-09-19', tags: 'github next localjev omlx bun typescript diffusiongemma 26b a4b json probabilities bakeoff m5 max benchmark qwen gemma moe wire compatible promptcoding' },
531
 
532
  /* ---------------- explain ---------------- */
533
  { cat: 'explain', title: '45-second TL;DR on Jev (video)', by: { name: 'Matija Sosic', handle: 'MatijaSosic' },
534
  desc: 'The most-shared explainer of the launch: a 45-second video that strips Jev down to its core idea, made because the official launch video was hard to follow.',
535
  quote: 'I find the core idea beautifully simple, but the video made it really hard to understand.',
536
  links: { tweet: 'https://x.com/MatijaSosic/status/2100190746389135772' },
537
- m: { likes: 8541, rts: 611, views: 1335770 }, date: '2026-09-16', tags: 'video tldr explainer launch' },
538
  { cat: 'explain', title: 'typesafe-ai-benchmark: an "imposter Jev" on Cerebras', by: { name: 'Shannon Duncan', handle: 'iamMrDuncan' },
539
  desc: 'An "imposter Jev" gateway on Cerebras with Qwen 3.8 27B, compared side by side. Similar quality and speed; TypeSafe far cheaper.',
540
  quote: 'Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost.',
541
  links: { tweet: 'https://x.com/iamMrDuncan/status/2100467548298899918', gh: 'https://github.com/iammrduncan/typesafe-ai-benchmark' },
542
- m: { likes: 208, rts: 25, views: 25518, stars: 30 }, date: '2026-09-16', tags: 'cerebras qwen gateway benchmark comparison cost' },
543
  { cat: 'explain', title: 'awesome-jev lists (AnotiaWang · hellogumbo)', by: { name: 'AnotiaWang · hellogumbo', handle: null },
544
  desc: 'Two curated indexes; AnotiaWang\'s "Research & Open Models" section is the tight one, hellogumbo\'s catches the long tail.',
545
  links: { gh: 'https://github.com/AnotiaWang/awesome-jev', gh2: 'https://github.com/hellogumbo/awesome-jev' },
546
- m: { stars: 73 }, date: '2026-09-16', tags: 'awesome list index curated roundup' },
547
  { cat: 'explain', title: 'jev-benchmarks: probability-aware eval harness', by: { name: 'AbdelStark', handle: null },
548
  desc: 'Calibration, selective risk and latency evals for typed decision models, Jev or open.',
549
  links: { gh: 'https://github.com/AbdelStark/jev-benchmarks' },
550
- m: { stars: 6 }, date: '2026-09-17', tags: 'calibration selective risk eval harness' },
551
  { cat: 'explain', title: 'Thread: System 1 / System 2 agents and the open recreations', by: { name: 'hhkkmon', handle: 'hhkkmon' },
552
  desc: '14-part explainer framing agents as "expensive LLM calls pretending to be if statements", listing the open recreations at the end.',
553
  quote: 'Big models think. Small decision models react.',
554
  links: { tweet: 'https://x.com/hhkkmon/status/2100443314957038010' },
555
- m: { likes: 2, rts: 0, views: 100 }, date: '2026-09-17', tags: 'thread explainer system 1 system 2 agents' },
556
  { cat: 'explain', title: 'Long-form: "OpenJev is NOT Jev open-sourced"', by: { name: 'mygtmhire', handle: 'mygtmhire' },
557
  desc: 'Most thorough single post on TheoLeeCJ/openjev, including the warning that its probabilities are not calibrated confidence.',
558
  quote: 'State in → probabilities out → code acts. No conversation required.',
559
  links: { tweet: 'https://x.com/mygtmhire/status/2100666488625734103' },
560
- m: { likes: 1, rts: 0, views: 95 }, date: '2026-09-17', tags: 'explainer openjev routing calibration' },
561
  { cat: 'explain', title: 'jev-behavior-study: probing the closed model', by: { name: 'RINNECODER', handle: null },
562
  desc: 'Black-box behavioural study of the closed Jev 1.13.0 itself. Not a replica — it measures the thing everyone is trying to reproduce.',
563
  links: { gh: 'https://github.com/RINNECODER/jev-behavior-study' },
@@ -608,7 +609,7 @@ const likesOf = it => it.m?.likes ?? -1;
608
  brand-new entry nobody has reacted to yet cannot climb on its date alone.
609
  The credit therefore rewards "new AND real", never novelty by itself. */
610
  const W = { likes: 1, stars: 5, hfLikes: 8, views: 1/500 };
611
- const RECENCY = { credit: 5000, halfLifeDays: 1, ref: 400 };
612
  const BASE_DATE = '2026-09-19'; /* the snapshot the metrics were last refreshed against */
613
  const ageDays = it => Math.max(0, (new Date(BASE_DATE) - new Date(it.date)) / 864e5);
614
  const recency = it => {
@@ -690,7 +691,7 @@ function orderedCats() {
690
  function cardHTML(it, cat, idx, q) {
691
  const m = it.m || {};
692
  const metrics = [];
693
- const tr = trending(it); if (tr) metrics.push(`<span class="tr" title="trending = ♥X + 5×★GitHub + 8×🤗likes + views/500 + recency credit (up to 5000, scaled by engagement, halving daily)">▲ ${fmt(tr)}</span>`);
694
  if (m.likes != null) metrics.push(`<span title="likes on X">${ICON.heart}${fmt(m.likes)}</span>`);
695
  if (m.rts) metrics.push(`<span title="reposts">${ICON.rt}${fmt(m.rts)}</span>`);
696
  if (m.views != null) metrics.push(`<span title="views on X">${ICON.eye}${fmt(m.views)}</span>`);
 
315
  <div>
316
  <p><b>Missing something?</b> Open a PR against this Space's <code class="mono">index.html</code> (the data is one JS array at the bottom of the file) or ping <a href="https://huggingface.co/multimodalart" target="_blank" rel="noopener">@multimodalart</a>. Hugging Face models and Spaces count as first-class artifacts here, not just GitHub repos.</p>
317
  <p><b>Trending</b> = ♥ likes on X + 5 × GitHub stars + 8 × Hub likes + views ÷ 500. All categories share one grid, color-coded; use the category chips to filter. Promised-but-unreleased items always sort last.</p>
318
+ <p>Engagement numbers are a snapshot from public X post metadata on 2026-09-19 and will drift. GitHub stars and Hub likes were read the same day. Descriptions paraphrase each project's own README and launch post; corrections welcome.</p>
319
  <p>Visual identity and outlined Huggies from <a href="https://huggingface.co/spaces/Chunte/HFBA" target="_blank" rel="noopener">HF Huggiverse (Chunte/HFBA)</a>. Unofficial, community-maintained, not affiliated with TypeSafe AI.</p>
320
  </div>
321
  <img src="huggies/rocket.png" alt="">
 
348
  desc: 'Batches every JSON key at once and reads category probabilities from stock Qwen 2.5 1B logits, ~5× faster than generating JSON.',
349
  quote: 'They were building in stealth for 2 years, I was building in stealth for 2 hours… No new training required.',
350
  links: { tweet: 'https://x.com/harshagundal/status/2100044305536889015', hf: 'https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD', space: 'https://huggingface.co/spaces/drinkmoonshine/parallel-constrained-decoding' },
351
+ m: { likes: 5796, rts: 355, views: 1269064, hfLikes: 429 }, date: '2026-09-16', tags: 'qwen json mac m4 on-device forks epsilon3 shreyansh26 drinkmoonshine' },
352
  { cat: 'decode', title: 'openjev-sglang: Jev-compatible public API', by: { name: 'Eric Zhang', handle: 'ekzhang1' },
353
  desc: 'A drop-in /v1/systemone endpoint over Qwen3.6-35B-A3B. SGLang\'s radix cache keeps the shared prefill hot, so 64 parallel tasks finish in under a second. Prefill-only, no generation.',
354
  quote: 'It runs a comparable open model (Qwen3.6-35B-A3B), and just uses SGLang radix cache to preserve the prefill reuse — 64 tasks in <1s.',
355
  links: { tweet: 'https://x.com/ekzhang1/status/2100651678110515383', gh: 'https://github.com/ekzhang/openjev-sglang' },
356
+ m: { likes: 977, rts: 60, views: 139265, stars: 196 }, date: '2026-09-17', tags: 'sglang qwen3.6 moe api systemone radix cache' },
357
  { cat: 'decode', title: 'openjev + openjev.com', by: { name: 'Theo Lee', handle: 'theoleecj' },
358
  desc: 'Reads option logits on Qwen3.5-4B instead of generating text: 21 decisions in 1.02s vs 5.33s for JSON. CUDA plus a WebGPU browser demo.',
359
  quote: 'On the Jev waitlist? You can use not Jev (but close enough) right here.',
360
  links: { tweet: 'https://x.com/theoleecj/status/2100243609060856264', gh: 'https://github.com/TheoLeeCJ/openjev', demo: 'https://openjev.com' },
361
+ m: { likes: 10, rts: 0, views: 3291, stars: 1976 }, date: '2026-09-16', tags: 'webgpu browser rtx 3090 qwen3.5 logits mit' },
362
  { cat: 'decode', title: 'PocketJev: on-device iPhone visual decisions', by: { name: 'aqty', handle: 'aqty' },
363
  desc: 'Qwen3-VL running locally on an iPhone via MLX. Camera frame + question + 3 options → direct option logits in ~1s, no photos saved.',
364
  quote: 'iPhone上でQwen 2Bをローカル実行して、カメラ画像+質問+3択から文章生成なしで直接判定。',
365
  links: { tweet: 'https://x.com/aqty/status/2100549517171175807', gh: 'https://github.com/NullPo-jp/PocketJev' },
366
+ m: { likes: 6, rts: 4, views: 424, stars: 0 }, date: '2026-09-17', tags: 'ios iphone mlx vision qwen3-vl camera japan' },
367
  { cat: 'decode', title: 'jevmlx: Jev-style decisions for any MLX model', by: { name: 'Beni', handle: 'beni_il_' },
368
  desc: 'Typed schema fields broadcast across the KV cache, one batch row per field, all scored in a single forward pass on any MLX model.',
369
  quote: 'If you can run Qwen on your machine, it can behave very similarly to Jev.',
370
  links: { tweet: 'https://x.com/beni_il_/status/2100617387116568956', gh: 'https://github.com/bnsd55/jevmlx' },
371
+ m: { likes: 7, rts: 0, views: 1183, stars: 40 }, date: '2026-09-17', tags: 'mlx apple silicon kv cache hebrew' },
372
  { cat: 'decode', title: 'jev-on-a-laptop: the honest benchmark', by: { name: 'Richard Becker', handle: 'RichardBcker1' },
373
  desc: 'Qwen 2.5 1.5B/7B and Qwen3 8B on a laptop vs Jev\'s published cases: 7B hits 73.8% agreement to Jev\'s 86.6%, 0.4–2s per decision. Finds confidence does not reliably flag errors.',
374
  quote: 'All code, benchmarks, model-size comparisons and the full write-up (unofficial, no affiliation with TypeSafe).',
375
  links: { tweet: 'https://x.com/RichardBcker1/status/2100195276107530393', gh: 'https://github.com/rorshopping/jev-on-a-laptop', space: 'https://huggingface.co/spaces/rorshopping/parallel-constrained-decisions' },
376
+ m: { likes: 1, rts: 0, views: 67, stars: 16 }, date: '2026-09-16', tags: 'benchmark study agreement fraud apple silicon m5 writeup explainer' },
377
  { cat: 'decode', title: 'jevfire: one context, many decisions on CUDA', by: { name: 'Enrique Alcázar', handle: 'enriquealcazarg' },
378
  desc: 'Parallel decisions for CUDA LLMs behind a vLLM-style API, with game-agent examples. Claims 10×+ over regular inference, runs in the browser, accepts images.',
379
  quote: 'Here you have an open-source approach to the same problem… uses the same principles as JEV, is type-safe, runs in your browser, and accepts images.',
380
  links: { tweet: 'https://x.com/enriquealcazarg/status/2100267330022375858', gh: 'https://github.com/kikoncuo/jevfire', site: 'https://kikoncuo.github.io/jevfire/' },
381
+ m: { likes: 1, rts: 0, views: 250, stars: 8 }, date: '2026-09-16', tags: 'cuda vllm browser images games primeagen' },
382
  { cat: 'decode', title: 'LFM2.5-RLCD (350M & 2.6B): PCD on Liquid models', by: { name: 'notnotsamuel · monotykamary', handle: null },
383
  desc: 'Parallel constrained decoding on unchanged Liquid LFM2.5 weights, branching attention and convolution state. Inference-only: 8–63× faster, field accuracy still ~60%.',
384
  links: { hf: 'https://huggingface.co/notnotsamuel/LFM2.5-350M-RLCD', hf2: 'https://huggingface.co/monotykamary/LFM2.5-2.6B-RLCD' },
385
+ m: { hfLikes: 18 }, date: '2026-09-16', tags: 'liquid lfm hub weights renamed inference only pcd' },
386
  { cat: 'decode', title: 'parallelConstraintDecoding (Java + Python)', by: { name: 'Stephan Janssen', handle: 'Stephan007' },
387
  desc: 'A Java port of the decode trick alongside a Python reference.',
388
  quote: 'Here\'s a java version of Jev 🤩',
389
  links: { tweet: 'https://x.com/Stephan007/status/2100673219967557906', gh: 'https://github.com/stephanj/parallelConstraintDecoding' },
390
+ m: { likes: 3, rts: 0, views: 103, stars: 10 }, date: '2026-09-17', tags: 'java jvm port' },
391
  { cat: 'decode', title: 'system-one (snellingio): "System One Lite"', by: { name: 'snellingio', handle: null },
392
  desc: 'Choice / Score / Noul server over MLX Qwen3, mirroring TypeSafe\'s three primitives. No launch post found.',
393
  links: { gh: 'https://github.com/snellingio/system-one' },
394
+ m: { stars: 38 }, date: '2026-09-16', tags: 'mlx qwen3 choice score noul server' },
395
  { cat: 'decode', title: 'open-jev (daseinlabs): Gemma 3 4B on MLX', by: { name: 'Dasein Labs', handle: null },
396
  desc: 'Prefix-cache option scoring on a local Gemma 3 4B via MLX, a Doom demo, and a /v1/systemone shim. Only seen in roundups.',
397
  links: { gh: 'https://github.com/daseinlabs/open-jev' },
398
+ m: { stars: 54 }, date: '2026-09-17', tags: 'gemma mlx doom systemone shim' },
399
  { cat: 'decode', title: 'openvons (open-Jev, Japanese)', by: { name: 'genai-craft', handle: null },
400
  desc: 'Decision layer answering finite-option questions with probabilities, from text, images or Japanese voice commands.',
401
  links: { gh: 'https://github.com/genai-craft/openvons' },
402
+ m: { stars: 9 }, date: '2026-09-17', tags: 'japanese voice image multimodal' },
403
  { cat: 'decode', title: 'openjev (zhihz): bilingual local decisions', by: { name: 'zhihz', handle: null },
404
  desc: 'Local bilingual probability decisions from context, questions and candidate answers.',
405
  links: { gh: 'https://github.com/zhihz/openjev' },
406
+ m: { stars: 14 }, date: '2026-09-16', tags: 'chinese bilingual local' },
407
  { cat: 'decode', title: 'system-one (sgoedecke): batched single-token choice', by: { name: 'Sean Goedecke', handle: null },
408
  desc: 'Batched single-token choice inference for open models, wire-compatible with TypeSafe.',
409
  links: { gh: 'https://github.com/sgoedecke/system-one' },
410
+ m: { stars: 14 }, date: '2026-09-17', tags: 'compatible batched single token' },
411
+ { cat: 'decode', title: 'reflex: stock Qwen3.5 / Qwen3-VL behind /v1/systemone', by: { name: 'Kshetrajna Raghavan', handle: 'kshetrajna' },
412
+ desc: 'Runs Jev\'s shape on stock Qwen3.5: choice / score / noul in one batched pass, a /v1/systemone server, and a WebGPU demo on a 0.8B model. Temperature scaling gets Qwen3.5-4B to 0.039 ECE vs Jev\'s 0.031; a LoRA recipe trains honest probabilities on public data.',
413
+ quote: '"Wonder if we could build that?" is a pretty normal response to new tech at Shopify. Jev got me curious, so I built Reflex: a Qwen-based experiment in structured decisions + probabilities running on WebGPU.',
414
+ links: { tweet: 'https://x.com/kshetrajna/status/2100739853101195744', gh: 'https://github.com/kshetrajna12/reflex', demo: 'https://kshetrajna12.github.io/reflex/' },
415
+ m: { likes: 330, rts: 22, views: 138281, stars: 77 }, date: '2026-09-17', tags: 'systemone choice score noul qwen3-vl image shared state server webgpu browser demo lora calibration ece shopify mmlu' },
416
  { cat: 'decode', title: 'LitJev: Jev-schema decision layer for the Qwen family', by: { name: 'Zhengxu Yu', handle: 'yuzxfred' },
417
  desc: 'Shared prefill, forked KV/recurrent state, one cached branch per question, option logits read from lm_head on any Qwen checkpoint (default Qwen3.8-27B), behind the exact Jev /v1/systemone schema. Playground, MMLU-Pro eval, optional temperature calibration.',
418
  quote: 'Qwen3.8-27B,10 道 MMLU-Pro 一次完成,0.3 s。比让模型只生成答案快 2.9x。非官方,基于公开信息。',
419
  links: { tweet: 'https://x.com/yuzxfred/status/2100652136878981337', gh: 'https://github.com/zhengxuyu/litjev', blog: 'https://x.com/yuzxfred/status/2100896621173047367' },
420
+ m: { likes: 2, rts: 0, views: 148, stars: 19 }, date: '2026-09-17', tags: 'qwen qwen3.8 27b systemone schema compatible calibration mmlu-pro doom chess vision transformers chinese explainer writeup limitations' },
421
 
422
  /* ---------------- trained ---------------- */
423
  { cat: 'decode', title: 'snapjudge: TypeSafe-compatible decisions on a Mac, scored against Jev', by: { name: 'Michael Gross', handle: 'michagross' },
424
  desc: 'Choice / Score / Noul behind /v1/systemone on MLX Qwen3.5-family models, probabilities read straight from the logits. On TypeSafe\'s 20 public cases (373 decisions) Qwen3.8-27B reaches 81.2% agreement vs Jev\'s 87.7%: level on short texts, behind on invoices.',
425
  links: { tweet: 'https://x.com/michagross/status/2101062157290856604', gh: 'https://github.com/Micha0827/snapjudge' },
426
+ m: { stars: 3, likes: 1, rts: 0, views: 11 }, date: '2026-09-18', tags: 'mlx apple silicon qwen3.6 qwen3.8 systemone api benchmark agreement german runner game' },
427
  { cat: 'trained', title: 'Bespoke Nimble: open data, open model, open recipe', by: { name: 'Mahesh Sathiamoorthy', handle: 'madiator' },
428
  desc: 'The first fully open recipe: LoRA on Qwen3.5-9B with the data curation published too. Contrastive curation flips one fact to make negatives, so training needs no probabilities. 90.1% agreement with Jev vs 66.4% for base Qwen.',
429
  quote: 'Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.',
430
  links: { tweet: 'https://x.com/madiator/status/2100990591215783946', gh: 'https://github.com/bespokelabsai/nimble', hf: 'https://huggingface.co/bespokelabs/Bespoke-Nimble-9B' },
431
+ m: { likes: 1210, rts: 151, views: 415617, stars: 509, hfLikes: 59 }, date: '2026-09-18', tags: 'bespokelabs nimble qwen3.5 9b lora contrastive data curation open recipe synthetic distillation free parallel scorer 100ms h100 macbook' },
432
  { cat: 'trained', title: 'jevlike: train your own one-pass option scorer', by: { name: 'Vinny LaRouge', handle: 'vinnylarouge' },
433
  desc: 'A tiny byte encoder (or frozen HF encoder) plus an option head maps context + N options to one probability each. Doom, chess and Wikispeedia demos; parent of several forks.',
434
  quote: 'I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes.',
435
  links: { tweet: 'https://x.com/vinnylarouge/status/2100170846346097083', gh: 'https://github.com/vinnylarouge/jevlike', demo: 'https://x.com/vinnylarouge/status/2100281651930513460' },
436
+ m: { likes: 2018, rts: 168, views: 164627, stars: 999 }, date: '2026-09-16', tags: 'train scorer encoder doom chess wikispeedia rl' },
437
  { cat: 'trained', title: 'openjev (AlexWortega): Qwen3.5 + MLP head, real weights', by: { name: 'Alex Wortega', handle: 'justALEXWORTEGA' },
438
  desc: 'The only trained "openjev" checkpoint on the Hub: Qwen3.5-4B as an NLI cross-encoder, argmax entailment is the move. Plays Doom zero-shot from text and pixels; 35B-A3B heads included.',
439
  quote: 'Typesafe: pnewed 💨 Jev: liberated 🫡 I trained an MLP on top of qwen 4b and it works literally like JEV',
440
  links: { tweet: 'https://x.com/justALEXWORTEGA/status/2100341039986798930', hf: 'https://huggingface.co/AlexWortega/openjev', gh: 'https://github.com/mikesmullin/openjev' },
441
+ m: { likes: 774, rts: 47, views: 103797, hfLikes: 213, stars: 2 }, date: '2026-09-16', tags: 'nli cross-encoder mlp qwen3.5 doom vision weights hub' },
442
  { cat: 'trained', title: 'decider-2b: full fine-tune of Qwen3.5-2B-Base', by: { name: 'Mapika', handle: null },
443
  desc: 'Full fine-tune of Qwen3.5-2B-Base on 942k examples from 64 decision datasets, plus a vision variant. Weights on the Hub.',
444
  links: { gh: 'https://github.com/Mapika/decider', hf: 'https://huggingface.co/Mapika/decider-2b', hf2: 'https://huggingface.co/Mapika/decider-2b-vision' },
445
+ m: { stars: 96, hfLikes: 49 }, date: '2026-09-16', tags: 'fine-tune qwen3.5 2b calibrated vision gh200 hub weights' },
446
  { cat: 'trained', title: 'rlcd-modernbert-151m ("OpenJev Verdict")', by: { name: 'heman10x', handle: null },
447
  desc: 'Post-trained ModernBERT (151M) scoring 25 candidate slots, abstention included, in under 35ms. WebGPU playground, Apache-2.0.',
448
  links: { hf: 'https://huggingface.co/heman10x/rlcd-modernbert-151m', gh: 'https://github.com/Heman10x-NGU/Verdict-open-jev' },
449
+ m: { hfLikes: 12, stars: 25 }, date: '2026-09-17', tags: 'modernbert encoder webgpu abstention gliclass hub weights' },
450
  { cat: 'trained', title: 'open-jev (pngwn): Qwen3.5-4B decision scorer + demo Space', by: { name: 'pngwn', handle: 'evilpingwin' },
451
  desc: 'Qwen3.5-4B LoRA fine-tune that scores each candidate answer for a typed question, with temperature scaling so confidence tracks accuracy. Weights (v1, v2b) and a demo Space on the Hub.',
452
  quote: 'Since everyone is open-jevving. I set ML Intern (on hugging chat) to work. After a few experiments and a little guidance and we did ok.',
453
  links: { tweet: 'https://x.com/evilpingwin/status/2100721265107751121', space: 'https://huggingface.co/spaces/pngwn/open-jev', hf: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer-v2b', hf2: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer' },
454
+ m: { likes: 230, rts: 36, views: 13184, hfLikes: 10 }, date: '2026-09-17', tags: 'qwen3.5 scoring head temperature scaling calibration ml intern huggingchat hub weights space demo' },
455
  { cat: 'trained', title: 'jevbetter: a stronger jevlike', by: { name: 'olanotolu', handle: null },
456
  desc: 'jevlike with a fancier head: hashed n-gram encoder, rival-aware attention, temperature scaling. Benchmarked against jevlike.',
457
  links: { gh: 'https://github.com/olanotolu/jevbetter' },
458
+ m: { stars: 12 }, date: '2026-09-16', tags: 'n-gram attention temperature scaling calibration' },
459
  { cat: 'trained', title: 'system-one-open: Gemma 4 E2B / Gemma 3 270M + head', by: { name: 'mithalouni', handle: null },
460
  desc: 'Typed calibrated decisions in one pass on Gemma 4 E2B and Gemma 3 270M, trained on Modal.',
461
  links: { gh: 'https://github.com/mithalouni/system-one-open' },
462
+ m: { stars: 11 }, date: '2026-09-17', tags: 'gemma modal scoring head' },
463
  { cat: 'trained', title: 'qwen-rlcd: Qwen3.5-0.8B Choice/Score/Noul', by: { name: 'shamazharikh', handle: null },
464
  desc: 'Choice / Score / Noul on Qwen3.5-0.8B. Inference works; training marked on hold.',
465
  links: { gh: 'https://github.com/shamazharikh/qwen-rlcd' },
 
467
  { cat: 'trained', title: 'system-one-gemma: Gemma 3 270M scoring head', by: { name: 'akash-kamat', handle: null },
468
  desc: 'Gemma 3 270M with a scoring head; single forward pass, no text generation.',
469
  links: { gh: 'https://github.com/akash-kamat/system-one-gemma' },
470
+ m: { stars: 2 }, date: '2026-09-17', tags: 'gemma 270m tiny' },
471
  { cat: 'trained', title: 'Laya: 421M RLCD decision model + demo Space', by: { name: 'Convai Innovations', handle: null },
472
  desc: 'ModernBERT-large fine-tuned plus a decision head trained from scratch (421M total) with RLCD against proper scoring rules: choice/score/noul with calibrated probabilities, all questions in one pass. Weights and a 7-tab Gradio demo on the Hub.',
473
  links: { gh: 'https://github.com/NandhaKishorM/laya', hf: 'https://huggingface.co/convaiinnovations/laya', space: 'https://huggingface.co/spaces/convaiinnovations/laya-demo', blog: 'https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me', site: 'https://www.reddit.com/r/LocalLLaMA/comments/1wjieap/made_the_horizontal_opensource_model_for_jev_with/' },
474
+ m: { stars: 1410, hfLikes: 577 }, date: '2026-09-18', tags: 'modernbert rlcd calibrated choice score noul support triage guardrails rag moderation routing hub weights space demo github pypi reddit devto' },
475
  { cat: 'trained', title: 'System One Mini: 69M DistilBERT, five fixed decisions', by: { name: 'David Hatley', handle: null },
476
  desc: '69M DistilBERT with five fixed typed heads over synthetic software-diagnosis summaries. Its card says: not a reproduction of Jev or RLCD.',
477
  links: { hf: 'https://huggingface.co/DavidHatley/system-one-mini' },
 
479
  { cat: 'trained', title: 'NanoJev: 0.6B Qwen3 with a real decision head', by: { name: 'TianyuCodings', handle: null },
480
  desc: 'A 0.6B Qwen3 backbone with an actual trained head emitting Choice / Boolean / Score across 2–255 candidates in one forward pass. Full training pipeline, maze and snake runs, weights and data on the Hub. The one Chinese and Korean roundups are pushing hardest.',
481
  links: { gh: 'https://github.com/TianyuCodings/NanoJev', hf: 'https://huggingface.co/C-Tianyu/NanoJev' },
482
+ m: { stars: 975, hfLikes: 35 }, date: '2026-09-17', tags: 'nanojev 0.6b qwen3 decision head choice boolean score maze snake training pipeline weights hub china korea' },
483
  { cat: 'trained', title: 'mini-jev: a preregistered measurement of the trick', by: { name: 'r-ms', handle: null },
484
  desc: 'Not another clone: a preregistered study of letter-logit versus grammar-JSON decoding on frozen Qwen3-4B and CLINC150. An actual measurement, no weights trained.',
485
  links: { gh: 'https://github.com/r-ms/mini-jev' },
486
+ m: { stars: 24 }, date: '2026-09-17', tags: 'preregistered study letter logits grammar json clinc150 qwen3-4b measurement' },
487
 
488
  { cat: 'trained', promised: true, title: 'Archer Hume\'s open-weight Jev (promised)', by: { name: 'Archer Hume', handle: '4rcherhume' },
489
  desc: 'Announced alongside the architecture teardown: an open-weight version trained on the inferred design. Training in progress, nothing released yet.',
490
  quote: 'I decided I\'d release an open-weight version. BUT training takes time, so while we all wait I decided I\'d drop the sauce.',
491
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
492
+ m: { likes: 2242, rts: 212, views: 376500 }, date: '2026-09-17', tags: 'promised coming soon open-weight training archer' },
493
 
494
  /* ---------------- prior ---------------- */
495
  { cat: 'prior', title: 'GLiNER 2.5: "Weights are on Hugging Face."', by: { name: 'George (Fastino / GLiNER)', handle: 'george_onx' },
496
  desc: 'Points the launch thread at the GLiNER 2.5 collection: a pretrained zero-shot encoder-classifier already doing single-pass classification over caller-supplied labels.',
497
  quote: 'We\'ve been reading the comments. Weights are on Hugging Face.',
498
  links: { tweet: 'https://x.com/george_onx/status/2100293114808119379', hf: 'https://huggingface.co/collections/fastino/gliner25-models' },
499
+ m: { likes: 1584, rts: 94, views: 224123 }, date: '2026-09-16', tags: 'gliner fastino zero-shot encoder classifier collection' },
500
  { cat: 'prior', title: 'DeepMost sales-conversion RL model + papers', by: { name: 'Nandakishor M', handle: 'Nandakishorm1' },
501
  desc: 'Claims the idea shipped in 2025 papers (arXiv 2503.23303, 2510.01237) with a Hub model and dataset. PPO over embeddings; loosely related.',
502
  quote: 'Also the second work published in September 2025 was exactly the same one jev proposed now.',
503
  links: { tweet: 'https://x.com/Nandakishorm1/status/2100451360873611666', hf: 'https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning', paper: 'https://arxiv.org/abs/2510.01237' },
504
+ m: { likes: 1, rts: 0, views: 84, hfLikes: 132 }, date: '2026-09-17', tags: 'ppo embeddings arxiv prior art sales' },
505
 
506
  /* ---------------- arch ---------------- */
507
  { cat: 'explain', title: 'Jev\'s Architecture Unmasked', by: { name: 'Archer Hume', handle: '4rcherhume' },
508
  desc: 'Argues Jev is a causal transformer with shared-state encoding, per-question branches and direct probability readouts, likely on a sparse MoE backbone, inferred from latency probing.',
509
  quote: 'All the grifters are completely wrong about Jev\'s architecture so I decided I\'d release an open-weight version. BUT training takes time…',
510
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
511
+ m: { likes: 2242, rts: 212, views: 376501 }, date: '2026-09-17', tags: 'moe teardown speculation open-weight promise blog' },
512
 
513
  /* ---------------- diffusion ---------------- */
514
  { cat: 'diffusion', title: 'DiffusionGemma-as-Jev + vLLM structured-generation patch', by: { name: 'Matt Mastracci', handle: 'mmastrac' },
515
  desc: 'Adds a Jev-like structured mode for DiffusionGemma to vLLM (PR #57250) and runs live evals: not slower than Jev\'s API and roughly tied on quality.',
516
  quote: 'Is Jev faster than DiffusionGemma? No ❌ Is Jev smarter than DiffusionGemma? No ❌ (they\'re roughly tied!)',
517
  links: { tweet: 'https://x.com/mmastrac/status/2100626193943052784', pr: 'https://github.com/vllm-project/vllm/pull/57250' },
518
+ m: { likes: 1003, rts: 75, views: 326616 }, date: '2026-09-17', tags: 'diffusion gemma vllm dgx spark evals' },
519
  { cat: 'diffusion', title: 'razorback16/openjev + api.codiv.ai host', by: { name: 'razorback16', handle: null },
520
  desc: 'The same vLLM PR #57250 route, packaged as a Jev-wire /v1/systemone service: a DiffusionGemma 26B-A4B canvas with only the label slots left as noise, one read-only denoise step reads out the distribution. Public host at api.codiv.ai.',
521
  links: { gh: 'https://github.com/razorback16/openjev', site: 'https://api.codiv.ai/v1/systemone' },
522
+ m: { stars: 134 }, date: '2026-09-17', tags: 'diffusiongemma 26b a4b nvfp4 vllm 57250 hosted api codiv systemone denoise canvas' },
523
  { cat: 'diffusion', title: 'open-jev (JoshuaSP): typed JSON with DiffusionGemma', by: { name: 'Joshua SP', handle: null },
524
  desc: 'Typed JSON inference with DiffusionGemma, benchmarked against Every and Jev.',
525
  links: { gh: 'https://github.com/JoshuaSP/open-jev' },
526
+ m: { stars: 16 }, date: '2026-09-16', tags: 'diffusion gemma benchmark every' },
527
  { cat: 'diffusion', title: 'LocalJev (GitHub Next): oMLX Jev API + model bake-off', by: { name: 'GitHub Next', handle: 'GitHubNext' },
528
  desc: 'Official GitHub Next bridge, since not everyone there has Jev access: a Bun server putting oMLX DiffusionGemma behind /v1/systemone. Prompts for JSON probabilities instead of reading logits: wire-compatible, not equivalent. Ships a 1,200-request bake-off.',
529
  quote: 'Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man\'s Jev on top of omlx for local use.',
530
  links: { tweet: 'https://x.com/GitHubNext/status/2101193436816920798', gh: 'https://github.com/githubnext/localjev' },
531
+ m: { likes: 1327, rts: 168, views: 185193, stars: 449 }, date: '2026-09-19', tags: 'github next localjev omlx bun typescript diffusiongemma 26b a4b json probabilities bakeoff m5 max benchmark qwen gemma moe wire compatible promptcoding' },
532
 
533
  /* ---------------- explain ---------------- */
534
  { cat: 'explain', title: '45-second TL;DR on Jev (video)', by: { name: 'Matija Sosic', handle: 'MatijaSosic' },
535
  desc: 'The most-shared explainer of the launch: a 45-second video that strips Jev down to its core idea, made because the official launch video was hard to follow.',
536
  quote: 'I find the core idea beautifully simple, but the video made it really hard to understand.',
537
  links: { tweet: 'https://x.com/MatijaSosic/status/2100190746389135772' },
538
+ m: { likes: 9277, rts: 668, views: 1584798 }, date: '2026-09-16', tags: 'video tldr explainer launch' },
539
  { cat: 'explain', title: 'typesafe-ai-benchmark: an "imposter Jev" on Cerebras', by: { name: 'Shannon Duncan', handle: 'iamMrDuncan' },
540
  desc: 'An "imposter Jev" gateway on Cerebras with Qwen 3.8 27B, compared side by side. Similar quality and speed; TypeSafe far cheaper.',
541
  quote: 'Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost.',
542
  links: { tweet: 'https://x.com/iamMrDuncan/status/2100467548298899918', gh: 'https://github.com/iammrduncan/typesafe-ai-benchmark' },
543
+ m: { likes: 209, rts: 25, views: 25998, stars: 32 }, date: '2026-09-17', tags: 'cerebras qwen gateway benchmark comparison cost' },
544
  { cat: 'explain', title: 'awesome-jev lists (AnotiaWang · hellogumbo)', by: { name: 'AnotiaWang · hellogumbo', handle: null },
545
  desc: 'Two curated indexes; AnotiaWang\'s "Research & Open Models" section is the tight one, hellogumbo\'s catches the long tail.',
546
  links: { gh: 'https://github.com/AnotiaWang/awesome-jev', gh2: 'https://github.com/hellogumbo/awesome-jev' },
547
+ m: { stars: 155 }, date: '2026-09-16', tags: 'awesome list index curated roundup' },
548
  { cat: 'explain', title: 'jev-benchmarks: probability-aware eval harness', by: { name: 'AbdelStark', handle: null },
549
  desc: 'Calibration, selective risk and latency evals for typed decision models, Jev or open.',
550
  links: { gh: 'https://github.com/AbdelStark/jev-benchmarks' },
551
+ m: { stars: 11 }, date: '2026-09-17', tags: 'calibration selective risk eval harness' },
552
  { cat: 'explain', title: 'Thread: System 1 / System 2 agents and the open recreations', by: { name: 'hhkkmon', handle: 'hhkkmon' },
553
  desc: '14-part explainer framing agents as "expensive LLM calls pretending to be if statements", listing the open recreations at the end.',
554
  quote: 'Big models think. Small decision models react.',
555
  links: { tweet: 'https://x.com/hhkkmon/status/2100443314957038010' },
556
+ m: { likes: 2, rts: 0, views: 119 }, date: '2026-09-17', tags: 'thread explainer system 1 system 2 agents' },
557
  { cat: 'explain', title: 'Long-form: "OpenJev is NOT Jev open-sourced"', by: { name: 'mygtmhire', handle: 'mygtmhire' },
558
  desc: 'Most thorough single post on TheoLeeCJ/openjev, including the warning that its probabilities are not calibrated confidence.',
559
  quote: 'State in → probabilities out → code acts. No conversation required.',
560
  links: { tweet: 'https://x.com/mygtmhire/status/2100666488625734103' },
561
+ m: { likes: 2, rts: 0, views: 123 }, date: '2026-09-17', tags: 'explainer openjev routing calibration' },
562
  { cat: 'explain', title: 'jev-behavior-study: probing the closed model', by: { name: 'RINNECODER', handle: null },
563
  desc: 'Black-box behavioural study of the closed Jev 1.13.0 itself. Not a replica — it measures the thing everyone is trying to reproduce.',
564
  links: { gh: 'https://github.com/RINNECODER/jev-behavior-study' },
 
609
  brand-new entry nobody has reacted to yet cannot climb on its date alone.
610
  The credit therefore rewards "new AND real", never novelty by itself. */
611
  const W = { likes: 1, stars: 5, hfLikes: 8, views: 1/500 };
612
+ const RECENCY = { credit: 10000, halfLifeDays: 1, ref: 400 };
613
  const BASE_DATE = '2026-09-19'; /* the snapshot the metrics were last refreshed against */
614
  const ageDays = it => Math.max(0, (new Date(BASE_DATE) - new Date(it.date)) / 864e5);
615
  const recency = it => {
 
691
  function cardHTML(it, cat, idx, q) {
692
  const m = it.m || {};
693
  const metrics = [];
694
+ const tr = trending(it); if (tr) metrics.push(`<span class="tr" title="trending = ♥X + 5×★GitHub + 8×🤗likes + views/500 + recency credit (up to 10000, scaled by engagement, halving daily)">▲ ${fmt(tr)}</span>`);
695
  if (m.likes != null) metrics.push(`<span title="likes on X">${ICON.heart}${fmt(m.likes)}</span>`);
696
  if (m.rts) metrics.push(`<span title="reposts">${ICON.rt}${fmt(m.rts)}</span>`);
697
  if (m.views != null) metrics.push(`<span title="views on X">${ICON.eye}${fmt(m.views)}</span>`);
og.html CHANGED
@@ -250,117 +250,118 @@ const ITEMS = [
250
  desc: 'Batches every JSON key at once and reads category probabilities from stock Qwen 2.5 1B logits, ~5× faster than generating JSON.',
251
  quote: 'They were building in stealth for 2 years, I was building in stealth for 2 hours… No new training required.',
252
  links: { tweet: 'https://x.com/harshagundal/status/2100044305536889015', hf: 'https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD', space: 'https://huggingface.co/spaces/drinkmoonshine/parallel-constrained-decoding' },
253
- m: { likes: 5507, rts: 318, views: 1145212, hfLikes: 343 }, date: '2026-09-16', tags: 'qwen json mac m4 on-device forks epsilon3 shreyansh26 drinkmoonshine' },
254
  { cat: 'decode', title: 'openjev-sglang: Jev-compatible public API', by: { name: 'Eric Zhang', handle: 'ekzhang1' },
255
  desc: 'A drop-in /v1/systemone endpoint over Qwen3.6-35B-A3B. SGLang\'s radix cache keeps the shared prefill hot, so 64 parallel tasks finish in under a second. Prefill-only, no generation.',
256
  quote: 'It runs a comparable open model (Qwen3.6-35B-A3B), and just uses SGLang radix cache to preserve the prefill reuse — 64 tasks in <1s.',
257
  links: { tweet: 'https://x.com/ekzhang1/status/2100651678110515383', gh: 'https://github.com/ekzhang/openjev-sglang' },
258
- m: { likes: 763, rts: 42, views: 91918, stars: 115 }, date: '2026-09-17', tags: 'sglang qwen3.6 moe api systemone radix cache' },
259
  { cat: 'decode', title: 'openjev + openjev.com', by: { name: 'Theo Lee', handle: 'theoleecj' },
260
  desc: 'Reads option logits on Qwen3.5-4B instead of generating text: 21 decisions in 1.02s vs 5.33s for JSON. CUDA plus a WebGPU browser demo.',
261
  quote: 'On the Jev waitlist? You can use not Jev (but close enough) right here.',
262
  links: { tweet: 'https://x.com/theoleecj/status/2100243609060856264', gh: 'https://github.com/TheoLeeCJ/openjev', demo: 'https://openjev.com' },
263
- m: { likes: 10, rts: 0, views: 3043, stars: 1427 }, date: '2026-09-16', tags: 'webgpu browser rtx 3090 qwen3.5 logits mit' },
264
  { cat: 'decode', title: 'PocketJev: on-device iPhone visual decisions', by: { name: 'aqty', handle: 'aqty' },
265
  desc: 'Qwen3-VL running locally on an iPhone via MLX. Camera frame + question + 3 options → direct option logits in ~1s, no photos saved.',
266
  quote: 'iPhone上でQwen 2Bをローカル実行して、カメラ画像+質問+3択から文章生成なしで直接判定。',
267
  links: { tweet: 'https://x.com/aqty/status/2100549517171175807', gh: 'https://github.com/NullPo-jp/PocketJev' },
268
- m: { likes: 6, rts: 4, views: 351, stars: 0 }, date: '2026-09-17', tags: 'ios iphone mlx vision qwen3-vl camera japan' },
269
  { cat: 'decode', title: 'jevmlx: Jev-style decisions for any MLX model', by: { name: 'Beni', handle: 'beni_il_' },
270
  desc: 'Typed schema fields broadcast across the KV cache, one batch row per field, all scored in a single forward pass on any MLX model.',
271
  quote: 'If you can run Qwen on your machine, it can behave very similarly to Jev.',
272
  links: { tweet: 'https://x.com/beni_il_/status/2100617387116568956', gh: 'https://github.com/bnsd55/jevmlx' },
273
- m: { likes: 7, rts: 0, views: 1154, stars: 19 }, date: '2026-09-17', tags: 'mlx apple silicon kv cache hebrew' },
274
  { cat: 'decode', title: 'jev-on-a-laptop: the honest benchmark', by: { name: 'Richard Becker', handle: 'RichardBcker1' },
275
  desc: 'Qwen 2.5 1.5B/7B and Qwen3 8B on a laptop vs Jev\'s published cases: 7B hits 73.8% agreement to Jev\'s 86.6%, 0.4–2s per decision. Finds confidence does not reliably flag errors.',
276
  quote: 'All code, benchmarks, model-size comparisons and the full write-up (unofficial, no affiliation with TypeSafe).',
277
  links: { tweet: 'https://x.com/RichardBcker1/status/2100195276107530393', gh: 'https://github.com/rorshopping/jev-on-a-laptop', space: 'https://huggingface.co/spaces/rorshopping/parallel-constrained-decisions' },
278
- m: { likes: 1, rts: 0, views: 62, stars: 14 }, date: '2026-09-16', tags: 'benchmark study agreement fraud apple silicon m5 writeup explainer' },
279
  { cat: 'decode', title: 'jevfire: one context, many decisions on CUDA', by: { name: 'Enrique Alcázar', handle: 'enriquealcazarg' },
280
  desc: 'Parallel decisions for CUDA LLMs behind a vLLM-style API, with game-agent examples. Claims 10×+ over regular inference, runs in the browser, accepts images.',
281
  quote: 'Here you have an open-source approach to the same problem… uses the same principles as JEV, is type-safe, runs in your browser, and accepts images.',
282
  links: { tweet: 'https://x.com/enriquealcazarg/status/2100267330022375858', gh: 'https://github.com/kikoncuo/jevfire', site: 'https://kikoncuo.github.io/jevfire/' },
283
- m: { likes: 1, rts: 0, views: 246, stars: 4 }, date: '2026-09-16', tags: 'cuda vllm browser images games primeagen' },
284
  { cat: 'decode', title: 'LFM2.5-RLCD (350M & 2.6B): PCD on Liquid models', by: { name: 'notnotsamuel · monotykamary', handle: null },
285
  desc: 'Parallel constrained decoding on unchanged Liquid LFM2.5 weights, branching attention and convolution state. Inference-only: 8–63× faster, field accuracy still ~60%.',
286
  links: { hf: 'https://huggingface.co/notnotsamuel/LFM2.5-350M-RLCD', hf2: 'https://huggingface.co/monotykamary/LFM2.5-2.6B-RLCD' },
287
- m: { hfLikes: 7 }, date: '2026-09-16', tags: 'liquid lfm hub weights renamed inference only pcd' },
288
  { cat: 'decode', title: 'parallelConstraintDecoding (Java + Python)', by: { name: 'Stephan Janssen', handle: 'Stephan007' },
289
  desc: 'A Java port of the decode trick alongside a Python reference.',
290
  quote: 'Here\'s a java version of Jev 🤩',
291
  links: { tweet: 'https://x.com/Stephan007/status/2100673219967557906', gh: 'https://github.com/stephanj/parallelConstraintDecoding' },
292
- m: { likes: 3, rts: 0, views: 99, stars: 6 }, date: '2026-09-17', tags: 'java jvm port' },
293
  { cat: 'decode', title: 'system-one (snellingio): "System One Lite"', by: { name: 'snellingio', handle: null },
294
  desc: 'Choice / Score / Noul server over MLX Qwen3, mirroring TypeSafe\'s three primitives. No launch post found.',
295
  links: { gh: 'https://github.com/snellingio/system-one' },
296
- m: { stars: 28 }, date: '2026-09-16', tags: 'mlx qwen3 choice score noul server' },
297
  { cat: 'decode', title: 'open-jev (daseinlabs): Gemma 3 4B on MLX', by: { name: 'Dasein Labs', handle: null },
298
  desc: 'Prefix-cache option scoring on a local Gemma 3 4B via MLX, a Doom demo, and a /v1/systemone shim. Only seen in roundups.',
299
  links: { gh: 'https://github.com/daseinlabs/open-jev' },
300
- m: { stars: 25 }, date: '2026-09-17', tags: 'gemma mlx doom systemone shim' },
301
  { cat: 'decode', title: 'openvons (open-Jev, Japanese)', by: { name: 'genai-craft', handle: null },
302
  desc: 'Decision layer answering finite-option questions with probabilities, from text, images or Japanese voice commands.',
303
  links: { gh: 'https://github.com/genai-craft/openvons' },
304
- m: { stars: 7 }, date: '2026-09-17', tags: 'japanese voice image multimodal' },
305
  { cat: 'decode', title: 'openjev (zhihz): bilingual local decisions', by: { name: 'zhihz', handle: null },
306
  desc: 'Local bilingual probability decisions from context, questions and candidate answers.',
307
  links: { gh: 'https://github.com/zhihz/openjev' },
308
- m: { stars: 5 }, date: '2026-09-16', tags: 'chinese bilingual local' },
309
  { cat: 'decode', title: 'system-one (sgoedecke): batched single-token choice', by: { name: 'Sean Goedecke', handle: null },
310
  desc: 'Batched single-token choice inference for open models, wire-compatible with TypeSafe.',
311
  links: { gh: 'https://github.com/sgoedecke/system-one' },
312
- m: { stars: 2 }, date: '2026-09-17', tags: 'compatible batched single token' },
313
- { cat: 'decode', title: 'reflex: stock Qwen3.5 / Qwen3-VL behind /v1/systemone', by: { name: 'kshetrajna12', handle: null },
314
- desc: 'Runnable server for the "architecture unmasked" reconstruction: choice / noul / score over untouched Qwen3.5 and Qwen3-VL, shared state with isolated questions, optional image state. Still updating.',
315
- links: { gh: 'https://github.com/kshetrajna12/reflex' },
316
- m: { stars: 48 }, date: '2026-09-17', tags: 'systemone choice score noul midtraining qwen3-vl image shared state server' },
 
317
  { cat: 'decode', title: 'LitJev: Jev-schema decision layer for the Qwen family', by: { name: 'Zhengxu Yu', handle: 'yuzxfred' },
318
  desc: 'Shared prefill, forked KV/recurrent state, one cached branch per question, option logits read from lm_head on any Qwen checkpoint (default Qwen3.8-27B), behind the exact Jev /v1/systemone schema. Playground, MMLU-Pro eval, optional temperature calibration.',
319
  quote: 'Qwen3.8-27B,10 道 MMLU-Pro 一次完成,0.3 s。比让模型只生成答案快 2.9x。非官方,基于公开信息。',
320
  links: { tweet: 'https://x.com/yuzxfred/status/2100652136878981337', gh: 'https://github.com/zhengxuyu/litjev', blog: 'https://x.com/yuzxfred/status/2100896621173047367' },
321
- m: { likes: 2, rts: 0, views: 125, stars: 4 }, date: '2026-09-17', tags: 'qwen qwen3.8 27b systemone schema compatible calibration mmlu-pro doom chess vision transformers chinese explainer writeup limitations' },
322
 
323
  /* ---------------- trained ---------------- */
324
- { cat: 'decode', title: 'snapjudge: TypeSafe-compatible decisions on a Mac, scored against Jev', by: { name: 'Michael Gross', handle: null },
325
  desc: 'Choice / Score / Noul behind /v1/systemone on MLX Qwen3.5-family models, probabilities read straight from the logits. On TypeSafe\'s 20 public cases (373 decisions) Qwen3.8-27B reaches 81.2% agreement vs Jev\'s 87.7%: level on short texts, behind on invoices.',
326
- links: { gh: 'https://github.com/Micha0827/snapjudge' },
327
- m: { stars: 0 }, date: '2026-09-18', tags: 'mlx apple silicon qwen3.6 qwen3.8 systemone api benchmark agreement german runner game' },
328
  { cat: 'trained', title: 'Bespoke Nimble: open data, open model, open recipe', by: { name: 'Mahesh Sathiamoorthy', handle: 'madiator' },
329
  desc: 'The first fully open recipe: LoRA on Qwen3.5-9B with the data curation published too. Contrastive curation flips one fact to make negatives, so training needs no probabilities. 90.1% agreement with Jev vs 66.4% for base Qwen.',
330
  quote: 'Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.',
331
  links: { tweet: 'https://x.com/madiator/status/2100990591215783946', gh: 'https://github.com/bespokelabsai/nimble', hf: 'https://huggingface.co/bespokelabs/Bespoke-Nimble-9B' },
332
- m: { likes: 163, rts: 16, views: 8393, stars: 13, hfLikes: 6 }, date: '2026-09-18', tags: 'bespokelabs nimble qwen3.5 9b lora contrastive data curation open recipe synthetic distillation free parallel scorer 100ms h100 macbook' },
333
  { cat: 'trained', title: 'jevlike: train your own one-pass option scorer', by: { name: 'Vinny LaRouge', handle: 'vinnylarouge' },
334
  desc: 'A tiny byte encoder (or frozen HF encoder) plus an option head maps context + N options to one probability each. Doom, chess and Wikispeedia demos; parent of several forks.',
335
  quote: 'I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes.',
336
  links: { tweet: 'https://x.com/vinnylarouge/status/2100170846346097083', gh: 'https://github.com/vinnylarouge/jevlike', demo: 'https://x.com/vinnylarouge/status/2100281651930513460' },
337
- m: { likes: 2006, rts: 168, views: 161595, stars: 818 }, date: '2026-09-16', tags: 'train scorer encoder doom chess wikispeedia rl' },
338
  { cat: 'trained', title: 'openjev (AlexWortega): Qwen3.5 + MLP head, real weights', by: { name: 'Alex Wortega', handle: 'justALEXWORTEGA' },
339
  desc: 'The only trained "openjev" checkpoint on the Hub: Qwen3.5-4B as an NLI cross-encoder, argmax entailment is the move. Plays Doom zero-shot from text and pixels; 35B-A3B heads included.',
340
  quote: 'Typesafe: pnewed 💨 Jev: liberated 🫡 I trained an MLP on top of qwen 4b and it works literally like JEV',
341
  links: { tweet: 'https://x.com/justALEXWORTEGA/status/2100341039986798930', hf: 'https://huggingface.co/AlexWortega/openjev', gh: 'https://github.com/mikesmullin/openjev' },
342
- m: { likes: 770, rts: 46, views: 101737, hfLikes: 113, stars: 1 }, date: '2026-09-16', tags: 'nli cross-encoder mlp qwen3.5 doom vision weights hub' },
343
  { cat: 'trained', title: 'decider-2b: full fine-tune of Qwen3.5-2B-Base', by: { name: 'Mapika', handle: null },
344
  desc: 'Full fine-tune of Qwen3.5-2B-Base on 942k examples from 64 decision datasets, plus a vision variant. Weights on the Hub.',
345
  links: { gh: 'https://github.com/Mapika/decider', hf: 'https://huggingface.co/Mapika/decider-2b', hf2: 'https://huggingface.co/Mapika/decider-2b-vision' },
346
- m: { stars: 86, hfLikes: 38 }, date: '2026-09-16', tags: 'fine-tune qwen3.5 2b calibrated vision gh200 hub weights' },
347
  { cat: 'trained', title: 'rlcd-modernbert-151m ("OpenJev Verdict")', by: { name: 'heman10x', handle: null },
348
  desc: 'Post-trained ModernBERT (151M) scoring 25 candidate slots, abstention included, in under 35ms. WebGPU playground, Apache-2.0.',
349
  links: { hf: 'https://huggingface.co/heman10x/rlcd-modernbert-151m', gh: 'https://github.com/Heman10x-NGU/Verdict-open-jev' },
350
- m: { hfLikes: 1, stars: 2 }, date: '2026-09-17', tags: 'modernbert encoder webgpu abstention gliclass hub weights' },
351
  { cat: 'trained', title: 'open-jev (pngwn): Qwen3.5-4B decision scorer + demo Space', by: { name: 'pngwn', handle: 'evilpingwin' },
352
  desc: 'Qwen3.5-4B LoRA fine-tune that scores each candidate answer for a typed question, with temperature scaling so confidence tracks accuracy. Weights (v1, v2b) and a demo Space on the Hub.',
353
  quote: 'Since everyone is open-jevving. I set ML Intern (on hugging chat) to work. After a few experiments and a little guidance and we did ok.',
354
  links: { tweet: 'https://x.com/evilpingwin/status/2100721265107751121', space: 'https://huggingface.co/spaces/pngwn/open-jev', hf: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer-v2b', hf2: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer' },
355
- m: { likes: 144, rts: 24, views: 8907, hfLikes: 4 }, date: '2026-09-17', tags: 'qwen3.5 scoring head temperature scaling calibration ml intern huggingchat hub weights space demo' },
356
  { cat: 'trained', title: 'jevbetter: a stronger jevlike', by: { name: 'olanotolu', handle: null },
357
  desc: 'jevlike with a fancier head: hashed n-gram encoder, rival-aware attention, temperature scaling. Benchmarked against jevlike.',
358
  links: { gh: 'https://github.com/olanotolu/jevbetter' },
359
- m: { stars: 10 }, date: '2026-09-16', tags: 'n-gram attention temperature scaling calibration' },
360
  { cat: 'trained', title: 'system-one-open: Gemma 4 E2B / Gemma 3 270M + head', by: { name: 'mithalouni', handle: null },
361
  desc: 'Typed calibrated decisions in one pass on Gemma 4 E2B and Gemma 3 270M, trained on Modal.',
362
  links: { gh: 'https://github.com/mithalouni/system-one-open' },
363
- m: { stars: 4 }, date: '2026-09-17', tags: 'gemma modal scoring head' },
364
  { cat: 'trained', title: 'qwen-rlcd: Qwen3.5-0.8B Choice/Score/Noul', by: { name: 'shamazharikh', handle: null },
365
  desc: 'Choice / Score / Noul on Qwen3.5-0.8B. Inference works; training marked on hold.',
366
  links: { gh: 'https://github.com/shamazharikh/qwen-rlcd' },
@@ -368,11 +369,11 @@ const ITEMS = [
368
  { cat: 'trained', title: 'system-one-gemma: Gemma 3 270M scoring head', by: { name: 'akash-kamat', handle: null },
369
  desc: 'Gemma 3 270M with a scoring head; single forward pass, no text generation.',
370
  links: { gh: 'https://github.com/akash-kamat/system-one-gemma' },
371
- m: { stars: 0 }, date: '2026-09-17', tags: 'gemma 270m tiny' },
372
  { cat: 'trained', title: 'Laya: 421M RLCD decision model + demo Space', by: { name: 'Convai Innovations', handle: null },
373
  desc: 'ModernBERT-large fine-tuned plus a decision head trained from scratch (421M total) with RLCD against proper scoring rules: choice/score/noul with calibrated probabilities, all questions in one pass. Weights and a 7-tab Gradio demo on the Hub.',
374
  links: { gh: 'https://github.com/NandhaKishorM/laya', hf: 'https://huggingface.co/convaiinnovations/laya', space: 'https://huggingface.co/spaces/convaiinnovations/laya-demo', blog: 'https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me', site: 'https://www.reddit.com/r/LocalLLaMA/comments/1wjieap/made_the_horizontal_opensource_model_for_jev_with/' },
375
- m: { stars: 163, hfLikes: 108 }, date: '2026-09-18', tags: 'modernbert rlcd calibrated choice score noul support triage guardrails rag moderation routing hub weights space demo github pypi reddit devto' },
376
  { cat: 'trained', title: 'System One Mini: 69M DistilBERT, five fixed decisions', by: { name: 'David Hatley', handle: null },
377
  desc: '69M DistilBERT with five fixed typed heads over synthetic software-diagnosis summaries. Its card says: not a reproduction of Jev or RLCD.',
378
  links: { hf: 'https://huggingface.co/DavidHatley/system-one-mini' },
@@ -380,86 +381,86 @@ const ITEMS = [
380
  { cat: 'trained', title: 'NanoJev: 0.6B Qwen3 with a real decision head', by: { name: 'TianyuCodings', handle: null },
381
  desc: 'A 0.6B Qwen3 backbone with an actual trained head emitting Choice / Boolean / Score across 2–255 candidates in one forward pass. Full training pipeline, maze and snake runs, weights and data on the Hub. The one Chinese and Korean roundups are pushing hardest.',
382
  links: { gh: 'https://github.com/TianyuCodings/NanoJev', hf: 'https://huggingface.co/C-Tianyu/NanoJev' },
383
- m: { stars: 235, hfLikes: 7 }, date: '2026-09-17', tags: 'nanojev 0.6b qwen3 decision head choice boolean score maze snake training pipeline weights hub china korea' },
384
  { cat: 'trained', title: 'mini-jev: a preregistered measurement of the trick', by: { name: 'r-ms', handle: null },
385
  desc: 'Not another clone: a preregistered study of letter-logit versus grammar-JSON decoding on frozen Qwen3-4B and CLINC150. An actual measurement, no weights trained.',
386
  links: { gh: 'https://github.com/r-ms/mini-jev' },
387
- m: { stars: 13 }, date: '2026-09-17', tags: 'preregistered study letter logits grammar json clinc150 qwen3-4b measurement' },
388
 
389
  { cat: 'trained', promised: true, title: 'Archer Hume\'s open-weight Jev (promised)', by: { name: 'Archer Hume', handle: '4rcherhume' },
390
  desc: 'Announced alongside the architecture teardown: an open-weight version trained on the inferred design. Training in progress, nothing released yet.',
391
  quote: 'I decided I\'d release an open-weight version. BUT training takes time, so while we all wait I decided I\'d drop the sauce.',
392
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
393
- m: { likes: 1982, rts: 182, views: 297794 }, date: '2026-09-17', tags: 'promised coming soon open-weight training archer' },
394
 
395
  /* ---------------- prior ---------------- */
396
  { cat: 'prior', title: 'GLiNER 2.5: "Weights are on Hugging Face."', by: { name: 'George (Fastino / GLiNER)', handle: 'george_onx' },
397
  desc: 'Points the launch thread at the GLiNER 2.5 collection: a pretrained zero-shot encoder-classifier already doing single-pass classification over caller-supplied labels.',
398
  quote: 'We\'ve been reading the comments. Weights are on Hugging Face.',
399
  links: { tweet: 'https://x.com/george_onx/status/2100293114808119379', hf: 'https://huggingface.co/collections/fastino/gliner25-models' },
400
- m: { likes: 1565, rts: 91, views: 219615 }, date: '2026-09-16', tags: 'gliner fastino zero-shot encoder classifier collection' },
401
  { cat: 'prior', title: 'DeepMost sales-conversion RL model + papers', by: { name: 'Nandakishor M', handle: 'Nandakishorm1' },
402
  desc: 'Claims the idea shipped in 2025 papers (arXiv 2503.23303, 2510.01237) with a Hub model and dataset. PPO over embeddings; loosely related.',
403
  quote: 'Also the second work published in September 2025 was exactly the same one jev proposed now.',
404
  links: { tweet: 'https://x.com/Nandakishorm1/status/2100451360873611666', hf: 'https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning', paper: 'https://arxiv.org/abs/2510.01237' },
405
- m: { likes: 0, rts: 0, views: 22, hfLikes: 86 }, date: '2026-09-17', tags: 'ppo embeddings arxiv prior art sales' },
406
 
407
  /* ---------------- arch ---------------- */
408
  { cat: 'explain', title: 'Jev\'s Architecture Unmasked', by: { name: 'Archer Hume', handle: '4rcherhume' },
409
  desc: 'Argues Jev is a causal transformer with shared-state encoding, per-question branches and direct probability readouts, likely on a sparse MoE backbone, inferred from latency probing.',
410
  quote: 'All the grifters are completely wrong about Jev\'s architecture so I decided I\'d release an open-weight version. BUT training takes time…',
411
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
412
- m: { likes: 1982, rts: 182, views: 297794 }, date: '2026-09-17', tags: 'moe teardown speculation open-weight promise blog' },
413
 
414
  /* ---------------- diffusion ---------------- */
415
  { cat: 'diffusion', title: 'DiffusionGemma-as-Jev + vLLM structured-generation patch', by: { name: 'Matt Mastracci', handle: 'mmastrac' },
416
  desc: 'Adds a Jev-like structured mode for DiffusionGemma to vLLM (PR #57250) and runs live evals: not slower than Jev\'s API and roughly tied on quality.',
417
  quote: 'Is Jev faster than DiffusionGemma? No ❌ Is Jev smarter than DiffusionGemma? No ❌ (they\'re roughly tied!)',
418
  links: { tweet: 'https://x.com/mmastrac/status/2100626193943052784', pr: 'https://github.com/vllm-project/vllm/pull/57250' },
419
- m: { likes: 575, rts: 36, views: 36393 }, date: '2026-09-17', tags: 'diffusion gemma vllm dgx spark evals' },
420
  { cat: 'diffusion', title: 'razorback16/openjev + api.codiv.ai host', by: { name: 'razorback16', handle: null },
421
  desc: 'The same vLLM PR #57250 route, packaged as a Jev-wire /v1/systemone service: a DiffusionGemma 26B-A4B canvas with only the label slots left as noise, one read-only denoise step reads out the distribution. Public host at api.codiv.ai.',
422
  links: { gh: 'https://github.com/razorback16/openjev', site: 'https://api.codiv.ai/v1/systemone' },
423
- m: { stars: 11 }, date: '2026-09-17', tags: 'diffusiongemma 26b a4b nvfp4 vllm 57250 hosted api codiv systemone denoise canvas' },
424
  { cat: 'diffusion', title: 'open-jev (JoshuaSP): typed JSON with DiffusionGemma', by: { name: 'Joshua SP', handle: null },
425
  desc: 'Typed JSON inference with DiffusionGemma, benchmarked against Every and Jev.',
426
  links: { gh: 'https://github.com/JoshuaSP/open-jev' },
427
- m: { stars: 14 }, date: '2026-09-16', tags: 'diffusion gemma benchmark every' },
428
  { cat: 'diffusion', title: 'LocalJev (GitHub Next): oMLX Jev API + model bake-off', by: { name: 'GitHub Next', handle: 'GitHubNext' },
429
  desc: 'Official GitHub Next bridge, since not everyone there has Jev access: a Bun server putting oMLX DiffusionGemma behind /v1/systemone. Prompts for JSON probabilities instead of reading logits: wire-compatible, not equivalent. Ships a 1,200-request bake-off.',
430
  quote: 'Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man\'s Jev on top of omlx for local use.',
431
  links: { tweet: 'https://x.com/GitHubNext/status/2101193436816920798', gh: 'https://github.com/githubnext/localjev' },
432
- m: { likes: 1084, rts: 148, views: 142175, stars: 328 }, date: '2026-09-19', tags: 'github next localjev omlx bun typescript diffusiongemma 26b a4b json probabilities bakeoff m5 max benchmark qwen gemma moe wire compatible promptcoding' },
433
 
434
  /* ---------------- explain ---------------- */
435
  { cat: 'explain', title: '45-second TL;DR on Jev (video)', by: { name: 'Matija Sosic', handle: 'MatijaSosic' },
436
  desc: 'The most-shared explainer of the launch: a 45-second video that strips Jev down to its core idea, made because the official launch video was hard to follow.',
437
  quote: 'I find the core idea beautifully simple, but the video made it really hard to understand.',
438
  links: { tweet: 'https://x.com/MatijaSosic/status/2100190746389135772' },
439
- m: { likes: 8541, rts: 611, views: 1335770 }, date: '2026-09-16', tags: 'video tldr explainer launch' },
440
  { cat: 'explain', title: 'typesafe-ai-benchmark: an "imposter Jev" on Cerebras', by: { name: 'Shannon Duncan', handle: 'iamMrDuncan' },
441
  desc: 'An "imposter Jev" gateway on Cerebras with Qwen 3.8 27B, compared side by side. Similar quality and speed; TypeSafe far cheaper.',
442
  quote: 'Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost.',
443
  links: { tweet: 'https://x.com/iamMrDuncan/status/2100467548298899918', gh: 'https://github.com/iammrduncan/typesafe-ai-benchmark' },
444
- m: { likes: 208, rts: 25, views: 25518, stars: 30 }, date: '2026-09-16', tags: 'cerebras qwen gateway benchmark comparison cost' },
445
  { cat: 'explain', title: 'awesome-jev lists (AnotiaWang · hellogumbo)', by: { name: 'AnotiaWang · hellogumbo', handle: null },
446
  desc: 'Two curated indexes; AnotiaWang\'s "Research & Open Models" section is the tight one, hellogumbo\'s catches the long tail.',
447
  links: { gh: 'https://github.com/AnotiaWang/awesome-jev', gh2: 'https://github.com/hellogumbo/awesome-jev' },
448
- m: { stars: 73 }, date: '2026-09-16', tags: 'awesome list index curated roundup' },
449
  { cat: 'explain', title: 'jev-benchmarks: probability-aware eval harness', by: { name: 'AbdelStark', handle: null },
450
  desc: 'Calibration, selective risk and latency evals for typed decision models, Jev or open.',
451
  links: { gh: 'https://github.com/AbdelStark/jev-benchmarks' },
452
- m: { stars: 6 }, date: '2026-09-17', tags: 'calibration selective risk eval harness' },
453
  { cat: 'explain', title: 'Thread: System 1 / System 2 agents and the open recreations', by: { name: 'hhkkmon', handle: 'hhkkmon' },
454
  desc: '14-part explainer framing agents as "expensive LLM calls pretending to be if statements", listing the open recreations at the end.',
455
  quote: 'Big models think. Small decision models react.',
456
  links: { tweet: 'https://x.com/hhkkmon/status/2100443314957038010' },
457
- m: { likes: 2, rts: 0, views: 100 }, date: '2026-09-17', tags: 'thread explainer system 1 system 2 agents' },
458
  { cat: 'explain', title: 'Long-form: "OpenJev is NOT Jev open-sourced"', by: { name: 'mygtmhire', handle: 'mygtmhire' },
459
  desc: 'Most thorough single post on TheoLeeCJ/openjev, including the warning that its probabilities are not calibrated confidence.',
460
  quote: 'State in → probabilities out → code acts. No conversation required.',
461
  links: { tweet: 'https://x.com/mygtmhire/status/2100666488625734103' },
462
- m: { likes: 1, rts: 0, views: 95 }, date: '2026-09-17', tags: 'explainer openjev routing calibration' },
463
  { cat: 'explain', title: 'jev-behavior-study: probing the closed model', by: { name: 'RINNECODER', handle: null },
464
  desc: 'Black-box behavioural study of the closed Jev 1.13.0 itself. Not a replica — it measures the thing everyone is trying to reproduce.',
465
  links: { gh: 'https://github.com/RINNECODER/jev-behavior-study' },
@@ -510,7 +511,7 @@ const likesOf = it => it.m?.likes ?? -1;
510
  brand-new entry nobody has reacted to yet cannot climb on its date alone.
511
  The credit therefore rewards "new AND real", never novelty by itself. */
512
  const W = { likes: 1, stars: 5, hfLikes: 8, views: 1/500 };
513
- const RECENCY = { credit: 5000, halfLifeDays: 1, ref: 400 };
514
  const BASE_DATE = '2026-09-19'; /* the snapshot the metrics were last refreshed against */
515
  const ageDays = it => Math.max(0, (new Date(BASE_DATE) - new Date(it.date)) / 864e5);
516
  const recency = it => {
 
250
  desc: 'Batches every JSON key at once and reads category probabilities from stock Qwen 2.5 1B logits, ~5× faster than generating JSON.',
251
  quote: 'They were building in stealth for 2 years, I was building in stealth for 2 hours… No new training required.',
252
  links: { tweet: 'https://x.com/harshagundal/status/2100044305536889015', hf: 'https://huggingface.co/harshatheg/Qwen-2.5-1B-RLCD', space: 'https://huggingface.co/spaces/drinkmoonshine/parallel-constrained-decoding' },
253
+ m: { likes: 5796, rts: 355, views: 1269064, hfLikes: 429 }, date: '2026-09-16', tags: 'qwen json mac m4 on-device forks epsilon3 shreyansh26 drinkmoonshine' },
254
  { cat: 'decode', title: 'openjev-sglang: Jev-compatible public API', by: { name: 'Eric Zhang', handle: 'ekzhang1' },
255
  desc: 'A drop-in /v1/systemone endpoint over Qwen3.6-35B-A3B. SGLang\'s radix cache keeps the shared prefill hot, so 64 parallel tasks finish in under a second. Prefill-only, no generation.',
256
  quote: 'It runs a comparable open model (Qwen3.6-35B-A3B), and just uses SGLang radix cache to preserve the prefill reuse — 64 tasks in <1s.',
257
  links: { tweet: 'https://x.com/ekzhang1/status/2100651678110515383', gh: 'https://github.com/ekzhang/openjev-sglang' },
258
+ m: { likes: 977, rts: 60, views: 139265, stars: 196 }, date: '2026-09-17', tags: 'sglang qwen3.6 moe api systemone radix cache' },
259
  { cat: 'decode', title: 'openjev + openjev.com', by: { name: 'Theo Lee', handle: 'theoleecj' },
260
  desc: 'Reads option logits on Qwen3.5-4B instead of generating text: 21 decisions in 1.02s vs 5.33s for JSON. CUDA plus a WebGPU browser demo.',
261
  quote: 'On the Jev waitlist? You can use not Jev (but close enough) right here.',
262
  links: { tweet: 'https://x.com/theoleecj/status/2100243609060856264', gh: 'https://github.com/TheoLeeCJ/openjev', demo: 'https://openjev.com' },
263
+ m: { likes: 10, rts: 0, views: 3291, stars: 1976 }, date: '2026-09-16', tags: 'webgpu browser rtx 3090 qwen3.5 logits mit' },
264
  { cat: 'decode', title: 'PocketJev: on-device iPhone visual decisions', by: { name: 'aqty', handle: 'aqty' },
265
  desc: 'Qwen3-VL running locally on an iPhone via MLX. Camera frame + question + 3 options → direct option logits in ~1s, no photos saved.',
266
  quote: 'iPhone上でQwen 2Bをローカル実行して、カメラ画像+質問+3択から文章生成なしで直接判定。',
267
  links: { tweet: 'https://x.com/aqty/status/2100549517171175807', gh: 'https://github.com/NullPo-jp/PocketJev' },
268
+ m: { likes: 6, rts: 4, views: 424, stars: 0 }, date: '2026-09-17', tags: 'ios iphone mlx vision qwen3-vl camera japan' },
269
  { cat: 'decode', title: 'jevmlx: Jev-style decisions for any MLX model', by: { name: 'Beni', handle: 'beni_il_' },
270
  desc: 'Typed schema fields broadcast across the KV cache, one batch row per field, all scored in a single forward pass on any MLX model.',
271
  quote: 'If you can run Qwen on your machine, it can behave very similarly to Jev.',
272
  links: { tweet: 'https://x.com/beni_il_/status/2100617387116568956', gh: 'https://github.com/bnsd55/jevmlx' },
273
+ m: { likes: 7, rts: 0, views: 1183, stars: 40 }, date: '2026-09-17', tags: 'mlx apple silicon kv cache hebrew' },
274
  { cat: 'decode', title: 'jev-on-a-laptop: the honest benchmark', by: { name: 'Richard Becker', handle: 'RichardBcker1' },
275
  desc: 'Qwen 2.5 1.5B/7B and Qwen3 8B on a laptop vs Jev\'s published cases: 7B hits 73.8% agreement to Jev\'s 86.6%, 0.4–2s per decision. Finds confidence does not reliably flag errors.',
276
  quote: 'All code, benchmarks, model-size comparisons and the full write-up (unofficial, no affiliation with TypeSafe).',
277
  links: { tweet: 'https://x.com/RichardBcker1/status/2100195276107530393', gh: 'https://github.com/rorshopping/jev-on-a-laptop', space: 'https://huggingface.co/spaces/rorshopping/parallel-constrained-decisions' },
278
+ m: { likes: 1, rts: 0, views: 67, stars: 16 }, date: '2026-09-16', tags: 'benchmark study agreement fraud apple silicon m5 writeup explainer' },
279
  { cat: 'decode', title: 'jevfire: one context, many decisions on CUDA', by: { name: 'Enrique Alcázar', handle: 'enriquealcazarg' },
280
  desc: 'Parallel decisions for CUDA LLMs behind a vLLM-style API, with game-agent examples. Claims 10×+ over regular inference, runs in the browser, accepts images.',
281
  quote: 'Here you have an open-source approach to the same problem… uses the same principles as JEV, is type-safe, runs in your browser, and accepts images.',
282
  links: { tweet: 'https://x.com/enriquealcazarg/status/2100267330022375858', gh: 'https://github.com/kikoncuo/jevfire', site: 'https://kikoncuo.github.io/jevfire/' },
283
+ m: { likes: 1, rts: 0, views: 250, stars: 8 }, date: '2026-09-16', tags: 'cuda vllm browser images games primeagen' },
284
  { cat: 'decode', title: 'LFM2.5-RLCD (350M & 2.6B): PCD on Liquid models', by: { name: 'notnotsamuel · monotykamary', handle: null },
285
  desc: 'Parallel constrained decoding on unchanged Liquid LFM2.5 weights, branching attention and convolution state. Inference-only: 8–63× faster, field accuracy still ~60%.',
286
  links: { hf: 'https://huggingface.co/notnotsamuel/LFM2.5-350M-RLCD', hf2: 'https://huggingface.co/monotykamary/LFM2.5-2.6B-RLCD' },
287
+ m: { hfLikes: 18 }, date: '2026-09-16', tags: 'liquid lfm hub weights renamed inference only pcd' },
288
  { cat: 'decode', title: 'parallelConstraintDecoding (Java + Python)', by: { name: 'Stephan Janssen', handle: 'Stephan007' },
289
  desc: 'A Java port of the decode trick alongside a Python reference.',
290
  quote: 'Here\'s a java version of Jev 🤩',
291
  links: { tweet: 'https://x.com/Stephan007/status/2100673219967557906', gh: 'https://github.com/stephanj/parallelConstraintDecoding' },
292
+ m: { likes: 3, rts: 0, views: 103, stars: 10 }, date: '2026-09-17', tags: 'java jvm port' },
293
  { cat: 'decode', title: 'system-one (snellingio): "System One Lite"', by: { name: 'snellingio', handle: null },
294
  desc: 'Choice / Score / Noul server over MLX Qwen3, mirroring TypeSafe\'s three primitives. No launch post found.',
295
  links: { gh: 'https://github.com/snellingio/system-one' },
296
+ m: { stars: 38 }, date: '2026-09-16', tags: 'mlx qwen3 choice score noul server' },
297
  { cat: 'decode', title: 'open-jev (daseinlabs): Gemma 3 4B on MLX', by: { name: 'Dasein Labs', handle: null },
298
  desc: 'Prefix-cache option scoring on a local Gemma 3 4B via MLX, a Doom demo, and a /v1/systemone shim. Only seen in roundups.',
299
  links: { gh: 'https://github.com/daseinlabs/open-jev' },
300
+ m: { stars: 54 }, date: '2026-09-17', tags: 'gemma mlx doom systemone shim' },
301
  { cat: 'decode', title: 'openvons (open-Jev, Japanese)', by: { name: 'genai-craft', handle: null },
302
  desc: 'Decision layer answering finite-option questions with probabilities, from text, images or Japanese voice commands.',
303
  links: { gh: 'https://github.com/genai-craft/openvons' },
304
+ m: { stars: 9 }, date: '2026-09-17', tags: 'japanese voice image multimodal' },
305
  { cat: 'decode', title: 'openjev (zhihz): bilingual local decisions', by: { name: 'zhihz', handle: null },
306
  desc: 'Local bilingual probability decisions from context, questions and candidate answers.',
307
  links: { gh: 'https://github.com/zhihz/openjev' },
308
+ m: { stars: 14 }, date: '2026-09-16', tags: 'chinese bilingual local' },
309
  { cat: 'decode', title: 'system-one (sgoedecke): batched single-token choice', by: { name: 'Sean Goedecke', handle: null },
310
  desc: 'Batched single-token choice inference for open models, wire-compatible with TypeSafe.',
311
  links: { gh: 'https://github.com/sgoedecke/system-one' },
312
+ m: { stars: 14 }, date: '2026-09-17', tags: 'compatible batched single token' },
313
+ { cat: 'decode', title: 'reflex: stock Qwen3.5 / Qwen3-VL behind /v1/systemone', by: { name: 'Kshetrajna Raghavan', handle: 'kshetrajna' },
314
+ desc: 'Runs Jev\'s shape on stock Qwen3.5: choice / score / noul in one batched pass, a /v1/systemone server, and a WebGPU demo on a 0.8B model. Temperature scaling gets Qwen3.5-4B to 0.039 ECE vs Jev\'s 0.031; a LoRA recipe trains honest probabilities on public data.',
315
+ quote: '"Wonder if we could build that?" is a pretty normal response to new tech at Shopify. Jev got me curious, so I built Reflex: a Qwen-based experiment in structured decisions + probabilities running on WebGPU.',
316
+ links: { tweet: 'https://x.com/kshetrajna/status/2100739853101195744', gh: 'https://github.com/kshetrajna12/reflex', demo: 'https://kshetrajna12.github.io/reflex/' },
317
+ m: { likes: 330, rts: 22, views: 138281, stars: 77 }, date: '2026-09-17', tags: 'systemone choice score noul qwen3-vl image shared state server webgpu browser demo lora calibration ece shopify mmlu' },
318
  { cat: 'decode', title: 'LitJev: Jev-schema decision layer for the Qwen family', by: { name: 'Zhengxu Yu', handle: 'yuzxfred' },
319
  desc: 'Shared prefill, forked KV/recurrent state, one cached branch per question, option logits read from lm_head on any Qwen checkpoint (default Qwen3.8-27B), behind the exact Jev /v1/systemone schema. Playground, MMLU-Pro eval, optional temperature calibration.',
320
  quote: 'Qwen3.8-27B,10 道 MMLU-Pro 一次完成,0.3 s。比让模型只生成答案快 2.9x。非官方,基于公开信息。',
321
  links: { tweet: 'https://x.com/yuzxfred/status/2100652136878981337', gh: 'https://github.com/zhengxuyu/litjev', blog: 'https://x.com/yuzxfred/status/2100896621173047367' },
322
+ m: { likes: 2, rts: 0, views: 148, stars: 19 }, date: '2026-09-17', tags: 'qwen qwen3.8 27b systemone schema compatible calibration mmlu-pro doom chess vision transformers chinese explainer writeup limitations' },
323
 
324
  /* ---------------- trained ---------------- */
325
+ { cat: 'decode', title: 'snapjudge: TypeSafe-compatible decisions on a Mac, scored against Jev', by: { name: 'Michael Gross', handle: 'michagross' },
326
  desc: 'Choice / Score / Noul behind /v1/systemone on MLX Qwen3.5-family models, probabilities read straight from the logits. On TypeSafe\'s 20 public cases (373 decisions) Qwen3.8-27B reaches 81.2% agreement vs Jev\'s 87.7%: level on short texts, behind on invoices.',
327
+ links: { tweet: 'https://x.com/michagross/status/2101062157290856604', gh: 'https://github.com/Micha0827/snapjudge' },
328
+ m: { stars: 3, likes: 1, rts: 0, views: 11 }, date: '2026-09-18', tags: 'mlx apple silicon qwen3.6 qwen3.8 systemone api benchmark agreement german runner game' },
329
  { cat: 'trained', title: 'Bespoke Nimble: open data, open model, open recipe', by: { name: 'Mahesh Sathiamoorthy', handle: 'madiator' },
330
  desc: 'The first fully open recipe: LoRA on Qwen3.5-9B with the data curation published too. Contrastive curation flips one fact to make negatives, so training needs no probabilities. 90.1% agreement with Jev vs 66.4% for base Qwen.',
331
  quote: 'Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.',
332
  links: { tweet: 'https://x.com/madiator/status/2100990591215783946', gh: 'https://github.com/bespokelabsai/nimble', hf: 'https://huggingface.co/bespokelabs/Bespoke-Nimble-9B' },
333
+ m: { likes: 1210, rts: 151, views: 415617, stars: 509, hfLikes: 59 }, date: '2026-09-18', tags: 'bespokelabs nimble qwen3.5 9b lora contrastive data curation open recipe synthetic distillation free parallel scorer 100ms h100 macbook' },
334
  { cat: 'trained', title: 'jevlike: train your own one-pass option scorer', by: { name: 'Vinny LaRouge', handle: 'vinnylarouge' },
335
  desc: 'A tiny byte encoder (or frozen HF encoder) plus an option head maps context + N options to one probability each. Doom, chess and Wikispeedia demos; parent of several forks.',
336
  quote: 'I reverse-engineered a jev-like architecture given its type. You can find the repo here to train your own jevlikes.',
337
  links: { tweet: 'https://x.com/vinnylarouge/status/2100170846346097083', gh: 'https://github.com/vinnylarouge/jevlike', demo: 'https://x.com/vinnylarouge/status/2100281651930513460' },
338
+ m: { likes: 2018, rts: 168, views: 164627, stars: 999 }, date: '2026-09-16', tags: 'train scorer encoder doom chess wikispeedia rl' },
339
  { cat: 'trained', title: 'openjev (AlexWortega): Qwen3.5 + MLP head, real weights', by: { name: 'Alex Wortega', handle: 'justALEXWORTEGA' },
340
  desc: 'The only trained "openjev" checkpoint on the Hub: Qwen3.5-4B as an NLI cross-encoder, argmax entailment is the move. Plays Doom zero-shot from text and pixels; 35B-A3B heads included.',
341
  quote: 'Typesafe: pnewed 💨 Jev: liberated 🫡 I trained an MLP on top of qwen 4b and it works literally like JEV',
342
  links: { tweet: 'https://x.com/justALEXWORTEGA/status/2100341039986798930', hf: 'https://huggingface.co/AlexWortega/openjev', gh: 'https://github.com/mikesmullin/openjev' },
343
+ m: { likes: 774, rts: 47, views: 103797, hfLikes: 213, stars: 2 }, date: '2026-09-16', tags: 'nli cross-encoder mlp qwen3.5 doom vision weights hub' },
344
  { cat: 'trained', title: 'decider-2b: full fine-tune of Qwen3.5-2B-Base', by: { name: 'Mapika', handle: null },
345
  desc: 'Full fine-tune of Qwen3.5-2B-Base on 942k examples from 64 decision datasets, plus a vision variant. Weights on the Hub.',
346
  links: { gh: 'https://github.com/Mapika/decider', hf: 'https://huggingface.co/Mapika/decider-2b', hf2: 'https://huggingface.co/Mapika/decider-2b-vision' },
347
+ m: { stars: 96, hfLikes: 49 }, date: '2026-09-16', tags: 'fine-tune qwen3.5 2b calibrated vision gh200 hub weights' },
348
  { cat: 'trained', title: 'rlcd-modernbert-151m ("OpenJev Verdict")', by: { name: 'heman10x', handle: null },
349
  desc: 'Post-trained ModernBERT (151M) scoring 25 candidate slots, abstention included, in under 35ms. WebGPU playground, Apache-2.0.',
350
  links: { hf: 'https://huggingface.co/heman10x/rlcd-modernbert-151m', gh: 'https://github.com/Heman10x-NGU/Verdict-open-jev' },
351
+ m: { hfLikes: 12, stars: 25 }, date: '2026-09-17', tags: 'modernbert encoder webgpu abstention gliclass hub weights' },
352
  { cat: 'trained', title: 'open-jev (pngwn): Qwen3.5-4B decision scorer + demo Space', by: { name: 'pngwn', handle: 'evilpingwin' },
353
  desc: 'Qwen3.5-4B LoRA fine-tune that scores each candidate answer for a typed question, with temperature scaling so confidence tracks accuracy. Weights (v1, v2b) and a demo Space on the Hub.',
354
  quote: 'Since everyone is open-jevving. I set ML Intern (on hugging chat) to work. After a few experiments and a little guidance and we did ok.',
355
  links: { tweet: 'https://x.com/evilpingwin/status/2100721265107751121', space: 'https://huggingface.co/spaces/pngwn/open-jev', hf: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer-v2b', hf2: 'https://huggingface.co/pngwn/system-one-qwen3.5-4b-scorer' },
356
+ m: { likes: 230, rts: 36, views: 13184, hfLikes: 10 }, date: '2026-09-17', tags: 'qwen3.5 scoring head temperature scaling calibration ml intern huggingchat hub weights space demo' },
357
  { cat: 'trained', title: 'jevbetter: a stronger jevlike', by: { name: 'olanotolu', handle: null },
358
  desc: 'jevlike with a fancier head: hashed n-gram encoder, rival-aware attention, temperature scaling. Benchmarked against jevlike.',
359
  links: { gh: 'https://github.com/olanotolu/jevbetter' },
360
+ m: { stars: 12 }, date: '2026-09-16', tags: 'n-gram attention temperature scaling calibration' },
361
  { cat: 'trained', title: 'system-one-open: Gemma 4 E2B / Gemma 3 270M + head', by: { name: 'mithalouni', handle: null },
362
  desc: 'Typed calibrated decisions in one pass on Gemma 4 E2B and Gemma 3 270M, trained on Modal.',
363
  links: { gh: 'https://github.com/mithalouni/system-one-open' },
364
+ m: { stars: 11 }, date: '2026-09-17', tags: 'gemma modal scoring head' },
365
  { cat: 'trained', title: 'qwen-rlcd: Qwen3.5-0.8B Choice/Score/Noul', by: { name: 'shamazharikh', handle: null },
366
  desc: 'Choice / Score / Noul on Qwen3.5-0.8B. Inference works; training marked on hold.',
367
  links: { gh: 'https://github.com/shamazharikh/qwen-rlcd' },
 
369
  { cat: 'trained', title: 'system-one-gemma: Gemma 3 270M scoring head', by: { name: 'akash-kamat', handle: null },
370
  desc: 'Gemma 3 270M with a scoring head; single forward pass, no text generation.',
371
  links: { gh: 'https://github.com/akash-kamat/system-one-gemma' },
372
+ m: { stars: 2 }, date: '2026-09-17', tags: 'gemma 270m tiny' },
373
  { cat: 'trained', title: 'Laya: 421M RLCD decision model + demo Space', by: { name: 'Convai Innovations', handle: null },
374
  desc: 'ModernBERT-large fine-tuned plus a decision head trained from scratch (421M total) with RLCD against proper scoring rules: choice/score/noul with calibrated probabilities, all questions in one pass. Weights and a 7-tab Gradio demo on the Hub.',
375
  links: { gh: 'https://github.com/NandhaKishorM/laya', hf: 'https://huggingface.co/convaiinnovations/laya', space: 'https://huggingface.co/spaces/convaiinnovations/laya-demo', blog: 'https://dev.to/nandakishor_m_6cc0adfde9f/i-built-non-autoregressive-decision-models-a-year-ago-then-a-frontier-lab-called-it-a-18me', site: 'https://www.reddit.com/r/LocalLLaMA/comments/1wjieap/made_the_horizontal_opensource_model_for_jev_with/' },
376
+ m: { stars: 1410, hfLikes: 577 }, date: '2026-09-18', tags: 'modernbert rlcd calibrated choice score noul support triage guardrails rag moderation routing hub weights space demo github pypi reddit devto' },
377
  { cat: 'trained', title: 'System One Mini: 69M DistilBERT, five fixed decisions', by: { name: 'David Hatley', handle: null },
378
  desc: '69M DistilBERT with five fixed typed heads over synthetic software-diagnosis summaries. Its card says: not a reproduction of Jev or RLCD.',
379
  links: { hf: 'https://huggingface.co/DavidHatley/system-one-mini' },
 
381
  { cat: 'trained', title: 'NanoJev: 0.6B Qwen3 with a real decision head', by: { name: 'TianyuCodings', handle: null },
382
  desc: 'A 0.6B Qwen3 backbone with an actual trained head emitting Choice / Boolean / Score across 2–255 candidates in one forward pass. Full training pipeline, maze and snake runs, weights and data on the Hub. The one Chinese and Korean roundups are pushing hardest.',
383
  links: { gh: 'https://github.com/TianyuCodings/NanoJev', hf: 'https://huggingface.co/C-Tianyu/NanoJev' },
384
+ m: { stars: 975, hfLikes: 35 }, date: '2026-09-17', tags: 'nanojev 0.6b qwen3 decision head choice boolean score maze snake training pipeline weights hub china korea' },
385
  { cat: 'trained', title: 'mini-jev: a preregistered measurement of the trick', by: { name: 'r-ms', handle: null },
386
  desc: 'Not another clone: a preregistered study of letter-logit versus grammar-JSON decoding on frozen Qwen3-4B and CLINC150. An actual measurement, no weights trained.',
387
  links: { gh: 'https://github.com/r-ms/mini-jev' },
388
+ m: { stars: 24 }, date: '2026-09-17', tags: 'preregistered study letter logits grammar json clinc150 qwen3-4b measurement' },
389
 
390
  { cat: 'trained', promised: true, title: 'Archer Hume\'s open-weight Jev (promised)', by: { name: 'Archer Hume', handle: '4rcherhume' },
391
  desc: 'Announced alongside the architecture teardown: an open-weight version trained on the inferred design. Training in progress, nothing released yet.',
392
  quote: 'I decided I\'d release an open-weight version. BUT training takes time, so while we all wait I decided I\'d drop the sauce.',
393
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
394
+ m: { likes: 2242, rts: 212, views: 376500 }, date: '2026-09-17', tags: 'promised coming soon open-weight training archer' },
395
 
396
  /* ---------------- prior ---------------- */
397
  { cat: 'prior', title: 'GLiNER 2.5: "Weights are on Hugging Face."', by: { name: 'George (Fastino / GLiNER)', handle: 'george_onx' },
398
  desc: 'Points the launch thread at the GLiNER 2.5 collection: a pretrained zero-shot encoder-classifier already doing single-pass classification over caller-supplied labels.',
399
  quote: 'We\'ve been reading the comments. Weights are on Hugging Face.',
400
  links: { tweet: 'https://x.com/george_onx/status/2100293114808119379', hf: 'https://huggingface.co/collections/fastino/gliner25-models' },
401
+ m: { likes: 1584, rts: 94, views: 224123 }, date: '2026-09-16', tags: 'gliner fastino zero-shot encoder classifier collection' },
402
  { cat: 'prior', title: 'DeepMost sales-conversion RL model + papers', by: { name: 'Nandakishor M', handle: 'Nandakishorm1' },
403
  desc: 'Claims the idea shipped in 2025 papers (arXiv 2503.23303, 2510.01237) with a Hub model and dataset. PPO over embeddings; loosely related.',
404
  quote: 'Also the second work published in September 2025 was exactly the same one jev proposed now.',
405
  links: { tweet: 'https://x.com/Nandakishorm1/status/2100451360873611666', hf: 'https://huggingface.co/DeepMostInnovations/sales-conversion-model-reinf-learning', paper: 'https://arxiv.org/abs/2510.01237' },
406
+ m: { likes: 1, rts: 0, views: 84, hfLikes: 132 }, date: '2026-09-17', tags: 'ppo embeddings arxiv prior art sales' },
407
 
408
  /* ---------------- arch ---------------- */
409
  { cat: 'explain', title: 'Jev\'s Architecture Unmasked', by: { name: 'Archer Hume', handle: '4rcherhume' },
410
  desc: 'Argues Jev is a causal transformer with shared-state encoding, per-question branches and direct probability readouts, likely on a sparse MoE backbone, inferred from latency probing.',
411
  quote: 'All the grifters are completely wrong about Jev\'s architecture so I decided I\'d release an open-weight version. BUT training takes time…',
412
  links: { tweet: 'https://x.com/4rcherhume/status/2100555442061820286', blog: 'https://archerhume.com/posts/jevs-architecture-unmasked/' },
413
+ m: { likes: 2242, rts: 212, views: 376501 }, date: '2026-09-17', tags: 'moe teardown speculation open-weight promise blog' },
414
 
415
  /* ---------------- diffusion ---------------- */
416
  { cat: 'diffusion', title: 'DiffusionGemma-as-Jev + vLLM structured-generation patch', by: { name: 'Matt Mastracci', handle: 'mmastrac' },
417
  desc: 'Adds a Jev-like structured mode for DiffusionGemma to vLLM (PR #57250) and runs live evals: not slower than Jev\'s API and roughly tied on quality.',
418
  quote: 'Is Jev faster than DiffusionGemma? No ❌ Is Jev smarter than DiffusionGemma? No ❌ (they\'re roughly tied!)',
419
  links: { tweet: 'https://x.com/mmastrac/status/2100626193943052784', pr: 'https://github.com/vllm-project/vllm/pull/57250' },
420
+ m: { likes: 1003, rts: 75, views: 326616 }, date: '2026-09-17', tags: 'diffusion gemma vllm dgx spark evals' },
421
  { cat: 'diffusion', title: 'razorback16/openjev + api.codiv.ai host', by: { name: 'razorback16', handle: null },
422
  desc: 'The same vLLM PR #57250 route, packaged as a Jev-wire /v1/systemone service: a DiffusionGemma 26B-A4B canvas with only the label slots left as noise, one read-only denoise step reads out the distribution. Public host at api.codiv.ai.',
423
  links: { gh: 'https://github.com/razorback16/openjev', site: 'https://api.codiv.ai/v1/systemone' },
424
+ m: { stars: 134 }, date: '2026-09-17', tags: 'diffusiongemma 26b a4b nvfp4 vllm 57250 hosted api codiv systemone denoise canvas' },
425
  { cat: 'diffusion', title: 'open-jev (JoshuaSP): typed JSON with DiffusionGemma', by: { name: 'Joshua SP', handle: null },
426
  desc: 'Typed JSON inference with DiffusionGemma, benchmarked against Every and Jev.',
427
  links: { gh: 'https://github.com/JoshuaSP/open-jev' },
428
+ m: { stars: 16 }, date: '2026-09-16', tags: 'diffusion gemma benchmark every' },
429
  { cat: 'diffusion', title: 'LocalJev (GitHub Next): oMLX Jev API + model bake-off', by: { name: 'GitHub Next', handle: 'GitHubNext' },
430
  desc: 'Official GitHub Next bridge, since not everyone there has Jev access: a Bun server putting oMLX DiffusionGemma behind /v1/systemone. Prompts for JSON probabilities instead of reading logits: wire-compatible, not equivalent. Ships a 1,200-request bake-off.',
431
  quote: 'Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man\'s Jev on top of omlx for local use.',
432
  links: { tweet: 'https://x.com/GitHubNext/status/2101193436816920798', gh: 'https://github.com/githubnext/localjev' },
433
+ m: { likes: 1327, rts: 168, views: 185193, stars: 449 }, date: '2026-09-19', tags: 'github next localjev omlx bun typescript diffusiongemma 26b a4b json probabilities bakeoff m5 max benchmark qwen gemma moe wire compatible promptcoding' },
434
 
435
  /* ---------------- explain ---------------- */
436
  { cat: 'explain', title: '45-second TL;DR on Jev (video)', by: { name: 'Matija Sosic', handle: 'MatijaSosic' },
437
  desc: 'The most-shared explainer of the launch: a 45-second video that strips Jev down to its core idea, made because the official launch video was hard to follow.',
438
  quote: 'I find the core idea beautifully simple, but the video made it really hard to understand.',
439
  links: { tweet: 'https://x.com/MatijaSosic/status/2100190746389135772' },
440
+ m: { likes: 9277, rts: 668, views: 1584798 }, date: '2026-09-16', tags: 'video tldr explainer launch' },
441
  { cat: 'explain', title: 'typesafe-ai-benchmark: an "imposter Jev" on Cerebras', by: { name: 'Shannon Duncan', handle: 'iamMrDuncan' },
442
  desc: 'An "imposter Jev" gateway on Cerebras with Qwen 3.8 27B, compared side by side. Similar quality and speed; TypeSafe far cheaper.',
443
  quote: 'Built an alternative version of @typesafeai but on @cerebras with Qwen 3.8 27b. Similar quality, similar performance, but vastly different cost.',
444
  links: { tweet: 'https://x.com/iamMrDuncan/status/2100467548298899918', gh: 'https://github.com/iammrduncan/typesafe-ai-benchmark' },
445
+ m: { likes: 209, rts: 25, views: 25998, stars: 32 }, date: '2026-09-17', tags: 'cerebras qwen gateway benchmark comparison cost' },
446
  { cat: 'explain', title: 'awesome-jev lists (AnotiaWang · hellogumbo)', by: { name: 'AnotiaWang · hellogumbo', handle: null },
447
  desc: 'Two curated indexes; AnotiaWang\'s "Research & Open Models" section is the tight one, hellogumbo\'s catches the long tail.',
448
  links: { gh: 'https://github.com/AnotiaWang/awesome-jev', gh2: 'https://github.com/hellogumbo/awesome-jev' },
449
+ m: { stars: 155 }, date: '2026-09-16', tags: 'awesome list index curated roundup' },
450
  { cat: 'explain', title: 'jev-benchmarks: probability-aware eval harness', by: { name: 'AbdelStark', handle: null },
451
  desc: 'Calibration, selective risk and latency evals for typed decision models, Jev or open.',
452
  links: { gh: 'https://github.com/AbdelStark/jev-benchmarks' },
453
+ m: { stars: 11 }, date: '2026-09-17', tags: 'calibration selective risk eval harness' },
454
  { cat: 'explain', title: 'Thread: System 1 / System 2 agents and the open recreations', by: { name: 'hhkkmon', handle: 'hhkkmon' },
455
  desc: '14-part explainer framing agents as "expensive LLM calls pretending to be if statements", listing the open recreations at the end.',
456
  quote: 'Big models think. Small decision models react.',
457
  links: { tweet: 'https://x.com/hhkkmon/status/2100443314957038010' },
458
+ m: { likes: 2, rts: 0, views: 119 }, date: '2026-09-17', tags: 'thread explainer system 1 system 2 agents' },
459
  { cat: 'explain', title: 'Long-form: "OpenJev is NOT Jev open-sourced"', by: { name: 'mygtmhire', handle: 'mygtmhire' },
460
  desc: 'Most thorough single post on TheoLeeCJ/openjev, including the warning that its probabilities are not calibrated confidence.',
461
  quote: 'State in → probabilities out → code acts. No conversation required.',
462
  links: { tweet: 'https://x.com/mygtmhire/status/2100666488625734103' },
463
+ m: { likes: 2, rts: 0, views: 123 }, date: '2026-09-17', tags: 'explainer openjev routing calibration' },
464
  { cat: 'explain', title: 'jev-behavior-study: probing the closed model', by: { name: 'RINNECODER', handle: null },
465
  desc: 'Black-box behavioural study of the closed Jev 1.13.0 itself. Not a replica — it measures the thing everyone is trying to reproduce.',
466
  links: { gh: 'https://github.com/RINNECODER/jev-behavior-study' },
 
511
  brand-new entry nobody has reacted to yet cannot climb on its date alone.
512
  The credit therefore rewards "new AND real", never novelty by itself. */
513
  const W = { likes: 1, stars: 5, hfLikes: 8, views: 1/500 };
514
+ const RECENCY = { credit: 10000, halfLifeDays: 1, ref: 400 };
515
  const BASE_DATE = '2026-09-19'; /* the snapshot the metrics were last refreshed against */
516
  const ageDays = it => Math.max(0, (new Date(BASE_DATE) - new Date(it.date)) / 864e5);
517
  const recency = it => {
og.png CHANGED

Git LFS Details

  • SHA256: f6fa43363b8b6f1bf28f5cfba8eac6240a718450cd0ba7d62f0dff3ad25cca20
  • Pointer size: 131 Bytes
  • Size of remote file: 362 kB

Git LFS Details

  • SHA256: 9a6648ac1653f2a62f20fdcdc46ba849f570443d28b1dc030e896410f09f5d0e
  • Pointer size: 131 Bytes
  • Size of remote file: 392 kB
refresh.js CHANGED
@@ -1,4 +1,4 @@
1
- const fs=require('fs');
2
  const P='index.html'; let src=fs.readFileSync(P,'utf8');
3
  const js=src.slice(src.indexOf('<script>')+8,src.indexOf('</script>'));
4
  const head=js.slice(0,js.indexOf('/* ---- state ---- */'));
@@ -6,6 +6,9 @@ const ITEMS=eval(head+';ITEMS');
6
  const sleep=ms=>new Promise(r=>setTimeout(r,ms));
7
  const clean=t=>t.split('').filter(c=>{const k=c.charCodeAt(0);return k>=32||c==='\n'||c==='\t'||c==='\r';}).join('');
8
  async function j(url,opt){const r=await fetch(url,opt);if(!r.ok)throw new Error(r.status+' '+url);return JSON.parse(clean(await r.text()));}
 
 
 
9
  const esc=t=>t.replace(/\\/g,'\\\\').replace(/'/g,"\\'");
10
  const changes=[];
11
  (async()=>{
@@ -14,18 +17,18 @@ const changes=[];
14
  if(L.tweet){const id=L.tweet.match(/status\/(\d+)/)[1];
15
  try{const d=await j('https://api.fxtwitter.com/status/'+id);const t=d.tweet;m.likes=t.likes;m.rts=t.retweets;m.views=t.views;}catch(e){notes.push('tweet:'+e.message);}}
16
  const ghs=[L.gh,L.gh2].filter(Boolean).map(u=>u.match(/github\.com\/([^\/]+\/[^\/#?]+)/)?.[1]).filter(Boolean);
17
- if(ghs.length){let sum=0,ok=true;for(const r of ghs){try{const d=await j('https://api.github.com/repos/'+r,{headers:{'User-Agent':'jev-tracker'}});sum+=d.stargazers_count;}catch(e){ok=false;notes.push('gh:'+e.message);}}if(ok)m.stars=sum;}
18
  const hfs=[L.hf,L.hf2].filter(Boolean).filter(u=>!u.includes('/collections/')&&!u.includes('/spaces/')).map(u=>u.match(/huggingface\.co\/([^\/]+\/[^\/#?]+)/)?.[1]).filter(Boolean);
19
- if(hfs.length){let sum=0,ok=true;for(const r of hfs){try{const d=await j('https://huggingface.co/api/models/'+r);sum+=d.likes||0;}catch(e){ok=false;notes.push('hf:'+e.message);}}if(ok)m.hfLikes=sum;}
20
  const before=JSON.stringify(it.m||{}),after=JSON.stringify(m);
21
- if(before!==after||notes.length)changes.push([it.title.slice(0,44),before,after,notes.join(' ')]);
22
  const ti=src.indexOf("title: '"+esc(it.title)+"'"); if(ti<0){console.log('TITLE NOT FOUND',it.title);continue;}
23
  const mi=src.indexOf('m: {',ti); const me=src.indexOf('}',mi)+1;
24
  const lit='m: { '+Object.entries(m).map(([k,v])=>k+': '+v).join(', ')+' }';
25
  src=src.slice(0,mi)+lit+src.slice(me);
26
- await sleep(150);
27
  }
28
  fs.writeFileSync(P,src);
29
- for(const c of changes)console.log(c.join(' | '));
30
  console.log('changed',changes.length,'of',ITEMS.length);
31
  })();
 
1
+ const fs=require('fs'),{execFileSync}=require('child_process');
2
  const P='index.html'; let src=fs.readFileSync(P,'utf8');
3
  const js=src.slice(src.indexOf('<script>')+8,src.indexOf('</script>'));
4
  const head=js.slice(0,js.indexOf('/* ---- state ---- */'));
 
6
  const sleep=ms=>new Promise(r=>setTimeout(r,ms));
7
  const clean=t=>t.split('').filter(c=>{const k=c.charCodeAt(0);return k>=32||c==='\n'||c==='\t'||c==='\r';}).join('');
8
  async function j(url,opt){const r=await fetch(url,opt);if(!r.ok)throw new Error(r.status+' '+url);return JSON.parse(clean(await r.text()));}
9
+ /* GitHub stars go through the gh CLI so the whole run fits in one authenticated
10
+ hour (5,000 requests); the unauthenticated API allows only 60. */
11
+ const ghStars=repo=>{try{return parseInt(execFileSync('gh',['api','repos/'+repo,'--jq','.stargazers_count'],{encoding:'utf8',stdio:['ignore','pipe','ignore']}).trim(),10);}catch(e){return null;}};
12
  const esc=t=>t.replace(/\\/g,'\\\\').replace(/'/g,"\\'");
13
  const changes=[];
14
  (async()=>{
 
17
  if(L.tweet){const id=L.tweet.match(/status\/(\d+)/)[1];
18
  try{const d=await j('https://api.fxtwitter.com/status/'+id);const t=d.tweet;m.likes=t.likes;m.rts=t.retweets;m.views=t.views;}catch(e){notes.push('tweet:'+e.message);}}
19
  const ghs=[L.gh,L.gh2].filter(Boolean).map(u=>u.match(/github\.com\/([^\/]+\/[^\/#?]+)/)?.[1]).filter(Boolean);
20
+ if(ghs.length){let sum=0,ok=true;for(const r of ghs){const s=ghStars(r);if(s===null){ok=false;notes.push('gh:'+r);}else sum+=s;}if(ok)m.stars=sum;}
21
  const hfs=[L.hf,L.hf2].filter(Boolean).filter(u=>!u.includes('/collections/')&&!u.includes('/spaces/')).map(u=>u.match(/huggingface\.co\/([^\/]+\/[^\/#?]+)/)?.[1]).filter(Boolean);
22
+ if(hfs.length){let sum=0,ok=true;for(const r of hfs){try{const d=await j('https://huggingface.co/api/models/'+r);sum+=d.likes||0;}catch(e){ok=false;notes.push('hf:'+r);}}if(ok)m.hfLikes=sum;}
23
  const before=JSON.stringify(it.m||{}),after=JSON.stringify(m);
24
+ if(before!==after||notes.length)changes.push([it.title.slice(0,44).padEnd(44),before,'->',after,notes.join(' ')]);
25
  const ti=src.indexOf("title: '"+esc(it.title)+"'"); if(ti<0){console.log('TITLE NOT FOUND',it.title);continue;}
26
  const mi=src.indexOf('m: {',ti); const me=src.indexOf('}',mi)+1;
27
  const lit='m: { '+Object.entries(m).map(([k,v])=>k+': '+v).join(', ')+' }';
28
  src=src.slice(0,mi)+lit+src.slice(me);
29
+ await sleep(60);
30
  }
31
  fs.writeFileSync(P,src);
32
+ for(const c of changes)console.log(c.join(' '));
33
  console.log('changed',changes.length,'of',ITEMS.length);
34
  })();