forge: regenerate the model card

README.md

@@ -1,19 +1,19 @@
 ---
-license: mit
+license: apache-2.0
+thumbnail: https://huggingface.co/AtomicChat/Qwen3.5-9B-DFlash-GGUF/resolve/main/hero.png
 base_model:
 - z-lab/Qwen3.5-9B-DFlash
 base_model_relation: quantized
-quantized_by: AlexAtomic
+quantized_by: AtomicChat
 pipeline_tag: text-generation
 library_name: gguf
 tags:
 - atomic-chat
-- dflash
-- speculative-decoding
-- draft-model
-- qwen
+- qwen3.5
+- z-lab
 - gguf
 - llama.cpp
+- quantized
 ---
 
 <center>
@@ -26,49 +26,89 @@ tags:
 
 <br/>
 
-<img src="https://huggingface.co/AtomicChat/Qwen3.5-9B-DFlash-GGUF/resolve/main/hero_dflash.png" alt="DFlash" style="width:100%; max-width:820px; height:auto; margin-bottom:0.6em;"/>
+<img src="https://huggingface.co/AtomicChat/Qwen3.5-9B-DFlash-GGUF/resolve/main/hero.png" alt="Qwen3.5 9B Dflash" style="width:100%; max-width:100%; height:auto; margin-bottom:0.6em;"/>
 
 <div style="display:flex; justify-content:center; gap:0.5em;">
-<a href="https://huggingface.co/z-lab/Qwen3.5-9B-DFlash"><strong>Draft: z-lab/Qwen3.5-9B-DFlash</strong></a> &nbsp;·&nbsp; <a href="https://huggingface.co/Qwen/Qwen3.5-9B"><strong>Target: Qwen/Qwen3.5-9B</strong></a>
+<a href="https://huggingface.co/z-lab/Qwen3.5-9B-DFlash"><strong>Base model: z-lab/Qwen3.5-9B-DFlash</strong></a>
 </div>
 </center>
 
-**Qwen3.5 9B DFlash**, the DFlash speculative-decoding **draft** converted to GGUF by [Atomic Chat](https://atomic.chat). Built straight from [z-lab](https://huggingface.co/z-lab)'s original weights. Runs fully offline.
+**Qwen3.5 9B Dflash**, self-quantized to GGUF by [Atomic Chat](https://atomic.chat). Built straight from Z Lab's original weights with a per-tensor importance matrix, so this is not a repack of somebody else's files. Runs fully offline.
 
-## What this is
+## Highlights
 
-[DFlash](https://github.com/z-lab/dflash) is a speculative-decoding method that drafts a whole **block** of candidate tokens in a single forward pass using a lightweight block-diffusion model, instead of one token at a time. This repo is the **draft component only** — it does nothing on its own. You run it alongside the target model **`Qwen/Qwen3.5-9B`**, which verifies the drafted block and keeps the longest correct prefix. Output is identical to running the target alone, just faster.
+- **1.3B parameters**: the weights this repo quantizes.
+- **Context length**: 262,144 tokens (256K), as published by Z Lab.
+- **6 layers**: Dense decoder, hybrid sliding-window (4096) and global attention.
+- **Full imatrix ladder**: every quant is calibrated with an importance matrix.
 
 > [!NOTE]
-> These GGUFs are **converted from z-lab's original weights**, not a repack of someone else's GGUF. The draft attaches to any GGUF of the target model (Atomic, unsloth, bartowski, ...).
+> These GGUFs are **self-quantized from the original weights**, not a repack. The importance matrix keeps low-bit quants closer to the full-precision model.
 
-## Run in llama.cpp
+> [!IMPORTANT]
+> Always pass `--jinja` so the **Qwen3.5 9B Dflash chat template** is applied. Without it the model can emit malformed turns.
 
-Needs a build of [llama.cpp](https://github.com/ggml-org/llama.cpp) with DFlash speculative decoding (PR #22105). You supply the target as `-m` and this draft as `-md`:
+## Model Overview
 
-```bash
-./llama-server \
-    -m   Qwen3.5-9B.gguf \
-    -md  Qwen3.5-9B-DFlash.Q8_0.gguf \
-    --spec-type draft-dflash --spec-draft-n-max 15 \
-    -ngl 99 -fa on --jinja -c 8192
-```
-
-DFlash is trained for **non-thinking** generation — pass `enable_thinking=false` in the chat template for best acceptance.
+| Property | Value |
+|---|---|
+| Base model | `z-lab/Qwen3.5-9B-DFlash` |
+| Parameters | 1.3B |
+| Layers | 6 |
+| Sliding window | 4096 tokens |
+| Context length | 262,144 tokens (256K) |
+| Vocabulary | 248,320 |
+| Modalities | Text |
+| Architecture | Dense decoder, hybrid sliding-window (4096) and global attention, 32 attention heads over 8 KV heads, `DFlashDraftModel` |
+| This repo | GGUF quants (imatrix). Quants: `Q8_0` |
 
 ## Choosing a quant
 
 | Quant | Size | Notes |
 |---|---|---|
-| **`Q8_0`** | 1.38 GB | Recommended. Near-lossless draft head, small and fast to draft with. |
+| **`Q8_0`** | 1.4 GB | **Effectively lossless, reference quality.** |
+
+> [!TIP]
+> Pick the largest file that fits your (V)RAM with room for context. `Q8_0` is the sweet spot for most setups; `Q6_K` or `Q8_0` for maximum fidelity.
+
+## Get started
+
+Run Qwen3.5 9B Dflash locally with:
+
+- **[Atomic Chat](https://atomic.chat):** the easiest path. Open the app, search `AtomicChat/Qwen3.5-9B-DFlash-GGUF`, pick a quant, hit **Use this model**.
+- **llama.cpp:** `llama-server -hf AtomicChat/Qwen3.5-9B-DFlash-GGUF:Q8_0 --jinja -c 8192`
+- **Ollama:** `ollama run hf.co/AtomicChat/Qwen3.5-9B-DFlash-GGUF:Q8_0`
+- **LM Studio / Jan:** search the repo id, download any quant.
+
+## Best practices
 
-## Performance
+| Parameter | Value |
+|---|---|
+| sampling defaults | not stated |
 
-<img src="https://huggingface.co/AtomicChat/Qwen3.5-9B-DFlash-GGUF/resolve/main/speedup.png" alt="Qwen3.5-9B-DFlash DFlash speedup" style="width:100%; max-width:840px; margin:0.4em 0;"/>
+The base model card does not state sampling defaults.
+
+## Run in llama.cpp
+
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cmake llama.cpp -B llama.cpp/build -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
+cmake --build llama.cpp/build --config Release -j --target llama-cli llama-server
+```
+
+```bash
+./llama.cpp/build/bin/llama-server \
+    -hf AtomicChat/Qwen3.5-9B-DFlash-GGUF:Q8_0 \
+    --jinja -ngl 99 -c 8192 -fa on
+```
 
-z-lab report up to **6.17x** lossless acceleration on their reference stack (vLLM / SGLang / Transformers). In `llama.cpp` today the DFlash port is newer: in our tests **dense** targets get roughly **1.8x-2.8x** end-to-end on code generation, and acceptance climbs on larger targets and structured/code output. Acceptance and speedup depend on the target and the content, not on the quantization. Speedups shrink on free-form prose and on small-active MoE targets.
+## How these were made
 
+1. Download `z-lab/Qwen3.5-9B-DFlash` (original weights).
+2. Convert to f16 GGUF with [llama.cpp](https://github.com/ggml-org/llama.cpp).
+3. Build an importance matrix over our calibration corpus.
+4. Quantize the ladder with `--imatrix`.
 
 ## License
 
-Released by z-lab under the MIT license. Converted to GGUF by Atomic Chat. See the [DFlash paper](https://arxiv.org/abs/2602.06036) and [project page](https://github.com/z-lab/dflash).
+Original model by Z Lab, released under the Apache 2.0 license. Quantized by Atomic Chat.