Update README.md

README.md

@@ -32,23 +32,20 @@ A Gated-DeltaNet-aware mixed GGUF of official
 for [llama.cpp](https://github.com/ggml-org/llama.cpp), Ollama, LM Studio,
 jan, KoboldCpp, and other stock GGUF runtimes.
 
-This is a quantization of the stock Qwen checkpoint — not a finetune, and
-not a new codec. Ridge is a probed mix of types llama.cpp already ships,
+This is a quantization of the Qwen3.8-27B checkpoint. Ridge is a probed mix of types
 written for this architecture: 64 layers =
 16 × `(3 × GatedDeltaNet → FFN + 1 × GatedAttn → FFN)`. Generic `IQ2_XS`
-and Unsloth UD-IQ2 do not treat GDN state (`ssm_alpha` / `ssm_beta`) or
-the GDN mixers as first-class. That is the gap.
+and UD-IQ2 do not treat GDN state (`ssm_alpha` / `ssm_beta`) or
+the GDN mixers as first-class. We fixed that.
 
 Nothing was stripped to make the file fit. The native MTP draft head
 (`blk.64` / `nextn`) stays in the GGUF. Vision is a separate BF16
 `mmproj`.
 
-The name is the measured bitrate: **3.69 bpw, 11.73 GiB**. It is not a
-2-bit file, and it is not lossless.
-
-This card is about choosing the file and running it. The official
-capability writeup lives on the
-**[base model card](https://huggingface.co/Qwen/Qwen3.8-27B)**.
+> [!Note]
+> This card is about choosing the file and running it. The official
+> capability writeup lives on the
+> **[base model card](https://huggingface.co/Qwen/Qwen3.8-27B)**.
 
 ---
 
@@ -68,7 +65,7 @@ If you only want text, download the Ridge GGUF. Add the `mmproj` for image input
 
 These are practical **weight-size-based estimates**, not a VRAM benchmark.
 They assume a modest context and leave room for runtime and the KV cache.
-Image input adds the 0.87 GiB `mmproj`. The native 262k window — and the
+Image input adds the 0.87 GiB `mmproj`. The native 262k window and the
 1M YaRN extension — make KV the dominant cost and may need offload
 regardless of weight quant.
 
@@ -92,18 +89,6 @@ layer. GDN state is disproportionately sensitive to low-bit quantization,
 so Ridge holds that path high and spends the saved bits by dropping
 mid-stack FFN.
 
-| Role | Tensors | Type |
-|---|---|---|
-| Embeddings, `lm_head` | `token_embd`, `output.weight` | Q6_K |
-| GDN state | `ssm_alpha`, `ssm_beta`, `ssm_conv1d`, `ssm_a`, `ssm_dt`, `ssm_norm` | Q8_0 |
-| Norms | `attn_norm`, `post_attention_norm`, `output_norm`, `attn_q_norm`, `attn_k_norm` | Q8_0 |
-| GDN mixers (48 linear-attn layers) | `attn_qkv`, `attn_gate`, `ssm_out` | Q4_K |
-| Full attention (layers 3, 7, …, 63) | `attn_q`, `attn_k`, `attn_v`, `attn_output` | Q5_K |
-| FFN first / last 4 layers (`blk.0–3`, `blk.60–63`) | `ffn_{gate,up,down}` | IQ3_S |
-| FFN mid (bulk of the bits) | `ffn_{gate,up,down}` | IQ2_M |
-| MTP draft (`blk.64`; no imatrix) | `nextn.*`, `blk.64.ffn_*` | Q6_K |
-| Vision | separate file | BF16 |
-
 **The Gated-DeltaNet state path is Q8_0.** Mixers are Q4_K, not IQ2.
 That is the difference between this file and a flat 2-bit dump of the
 same model.
@@ -125,11 +110,6 @@ Same box, same calibration file, `llama-perplexity`, 80 chunks,
 | BF16 GGUF (this convert) | 50.89 GiB | 16.00 | **7.15 ± 0.12** | — |
 | **Ridge-3.7bpw** | **11.73 GiB** | **3.69** | **7.82 ± 0.14** | **+9.3 %** |
 
-Not lossless. +9 % PPL at 23 % of BF16 size is the first-bake number.
-A same-suite bake-off against Unsloth `UD-Q3_K_XL` and bartowski
-`IQ3_XXS` has not been run on this box yet — do not read the size table
-below as a quality ranking.
-
 ---
 
 ## Comparison
@@ -143,24 +123,10 @@ where we measured the file ourselves.
 | `UD-IQ2_XXS` | [unsloth](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF) | 8.39 GiB | ~2.1 bpw | *not measured here* (Unsloth quotes 82.5 % top-1 vs BF16) |
 | `UD-IQ2_M` | unsloth | 9.61 GiB | ~2.4 bpw | *not measured* |
 | `IQ2_XXS` | [bartowski](https://huggingface.co/bartowski/Qwen3.8-27B-GGUF) | 8.75 GiB | ~2.2 bpw | *not measured* |
-| `IQ2_M` | bartowski | 10.13 GiB | ~2.5 bpw | *not measured* |
-| `UD-IQ3_XXS` | unsloth | 11.10 GiB | ~2.8 bpw | *not measured* |
-| `Q2_K` | bartowski | 11.03 GiB | ~2.7 bpw | *not measured* |
 | `Q3_K_S` | unsloth | 11.71 GiB | ~3.1 bpw | *not measured* |
 | **Ridge-3.7bpw** | **empero-ai** | **11.73 GiB** | **3.69 bpw** | **7.82 (+9 %)** |
 | `IQ3_XXS` | bartowski | 11.76 GiB | ~2.9 bpw | *not measured* |
 | `UD-Q3_K_XL` | unsloth | 12.52 GiB | ~3.4 bpw | *not measured* |
-| `Q3_K_M` | unsloth | 12.87 GiB | ~3.5 bpw | *not measured* |
-| `Q4_K_M` | unsloth | 15.93 GiB | ~4.3 bpw | *not measured* |
-
-How to read this:
-
-- Ridge-3.7bpw sits in the **~12 GiB** band, next to Unsloth `Q3_K_S` and
-  bartowski `IQ3_XXS` — not next to UD-IQ2_XXS.
-- The extra bits versus a flat IQ2 go into **GDN state (Q8_0)**, **GDN
-  mixers (Q4_K)**, and **full attention (Q5_K)**. Mid-stack FFN stays IQ2_M.
-- We are not claiming this file beats Unsloth 2-bit at the same size. We
-  spent the bits on GDN on purpose.
 
 ---
 
@@ -288,12 +254,7 @@ YaRN. Set `-c` to what you actually need — the KV cache, not the
 
 ## Limitations
 
-- **Not lossless.** +9 % wiki-style PPL vs our BF16 convert is a real
-  gap. Do not describe this as an almost-lossless 2-bit.
-- **Not a 2-bit file.** The name is Ridge-**3.7bpw** because that is
-  what llama-quantize reported (3.69).
-- **Same-suite bake-off pending.** Unsloth / bartowski PPL on this
-  calibration set has not been measured here.
+- **Not lossless.** +9 % wiki-style PPL vs our BF16 convert
 - **Context costs memory.** Weight size is only part of the hardware
   budget.
 - **MTP is runtime-dependent.** The head is in the file; the speedup