Avifenesh/Hy3-REAP-Layer103p5-bw24

Hy3 REAP Layer103.5 for bw24

This repository contains the receipt-bound layer103-late20 expert overlay for bw24. It preserves the complete Layer100 allocation, then restores 262 experts only in layers 60-79. The result is a 103,489,802,752-byte logical model (96.38 GiB) represented by 78,490,288,128 bytes of expert payloads plus non-expert fallback tensors from the pinned Hy3 source.

Matched directional result

arm	math	code	history	other	total
Layer100 matched	27	32	3	11	73/115
Layer103.5 late20	29	32	4	11	76/115

The paired result is 5 wins, 2 losses, and 108 ties. The domain-macro bootstrap interval crosses zero and the paired exact-sign p-value is 0.453125, so this is directional evidence rather than a statistical-significance claim. Public capability scores were not used for construction or healing; the arm was selected from private routing displacement and layer position only.

Artifact contract

This is a bw24-expert-overlay-v2, not a generic Transformers, MLX, GGUF, or ModelOpt checkpoint. Use it with bw24's Hy3 overlay loader. The immutable published manifest points at the build machine's source path because that path is part of the evaluation receipt. At runtime, create a lightweight relocated view that retains the published expert files by symlink.

Required source:

• model: tencent/Hy3
• revision: 716aa7241bd6d95896be4ebfc761162a9c4d49ef
• 99 shards, 597,578,239,288 bytes
• config.json SHA-256: 663036ceca3d8a178cd772739566c262caffdecebaed6c1d76b464d729bb2951
• model.safetensors.index.json SHA-256: 9594f1a9419e62ca7afca51bb644f38ef19039374f7812449381ccf42f0ef79b

Download and run

hf download Avifenesh/Hy3-REAP-Layer103p5-bw24 --local-dir /models/hy3-layer103p5
hf download tencent/Hy3 --revision 716aa7241bd6d95896be4ebfc761162a9c4d49ef \
  --local-dir /models/hy3-source

python /path/to/bw24/tools/relocate_hy3_expert_overlay.py \
  /models/hy3-layer103p5 \
  /models/hy3-source \
  /models/hy3-layer103p5-runtime

BW24_MODELS=hy3-layer103p5=/models/hy3-layer103p5-runtime \
BW24_COMPAT=openai BW24_SERVE_SPEC=0 BW24_KV_REUSE=0 \
  /path/to/bw24-server

MTP/speculative decoding and KV reuse were disabled in the matched evaluation.

Receipt anchors

• research branch head at release preparation: 86ec0d138d4066318b49696abe74648b08f2ed67
• source plan: 9606b1b96890b270534237b1143a5f5f25165245d1b5f08f515c25268d1b056c
• artifact manifest: 08f206aed555752982585a59a7b5096b9cc6e71faf1f84ad5c6dd60476b7509a
• screen summary: 661e495c02467acf5f28180eb73d562d25fea8724449e69f305831beb435be3d
• winner receipt: 4c6d4089d5dec428a5575b8c9971521f1f04ed5041e48e02bda67d3dae993ab3
• bounded-screen receipt: b9a76d472f466e8d6b6a9d2e489eac889de03af9cf740c2094ad1f9c6cbcb1e2

The machine-readable plan, source receipt, winner receipt, and full screen summary are under evidence/. Per-file quantization metadata and payload offsets are in manifest.json and the *.complete.json files.

Related

• Engine: github.com/avifenesh/bw24 — the spill runbook lives in docs/HY3-SPILL.md.
• Avifenesh/bw24-bench — benchmark artifacts and trimmed speculative-decode drafts for the six board models.