---

license: other

license_name: qwen-community-1.0

license_link: LICENSE

base_model: Qwen/Qwen3.8-Flash-Next

pipeline_tag: text-generation

library_name: gguf

tags:

• gguf
• quantization
• experimental
• hadamard

---

Qwen3.8-Flash-Next Q2_0 H64-S4 GGUF

I wanted a smaller GGUF I could run, and I wanted to know more than just whether it loaded. This is where that experiment landed. How we got here has the story, including the first Q2 export, the H64-S4 experiment, the interrupted run, and what I could and couldn't verify.

You need a runtime that supports the matching Hadamard transform. A runtime that ignores the rotation metadata can compute the wrong result. The tested llama.cpp changes, source patch, and build instructions are in runtime/.

What's in the files

The export has nine GGUF shards, about 68.9 GB in total. It is based on Qwen/Qwen3.8-Flash-Next, revision de4b8e4d43b917e7706784d8bb445c9af86a3540.

I used mixed precision to spend the bits where the recipe calls for them:

• The 144 routed expert gate, up, and down matrices use Q2_0 ternary weights, fitted in groups of 64 with FP16 scales.
• Before fitting those expert weights, the exporter applies normalized Walsh-Hadamard transforms in blocks of 64. A fixed sign pattern from seed 4 makes the transform reproducible. The runtime applies the matching input transform.
• The output head, attention weights, and shared experts use Q8_0 where selected by the recipe. The per-layer n-gram table uses Q4_0. Protected control tensors retain their source types.
• The MTP head and vision projector are not included. The inference checks used MTP disabled.

This is post-training quantization. There was no recovery training, and this is not a trained Bonsai release.

How the test went

I tested the full model on three RTX 3090 GPUs, with the per-layer token table on CPU and a 65,536-token context allocation. The context setting shows what I configured; it is not a claim that answer quality was checked across the full context window.

In a small diagnostic run, H64-S4 passed 21 of 27 prompts with thinking off. All five generated code samples passed isolated execution checks. JSON formatting, Spanish, and both context-recall prompts also passed. Six short arithmetic and logic answers missed. When I retried those same six with thinking on, all six passed.

These prompts are a useful smoke check, not a broad model benchmark. The original full Q2 file had passed its own load and prompt checks, but I could not find it later to run the same prompts side by side. So I can't claim H64-S4 gives better answers than that Q2 export.

A separate weight-sample sweep found 2.91% lower normalized reconstruction error for H64-S4 than direct Q2 on the 26 samples used to choose the setting. That result helped me decide to try the full export; it does not prove a general accuracy gain.

See the original prompts and responses, the thinking-mode follow-up, and the tensor checks.

Download and run

Download all nine shards into the same folder and check them with sha256sum -c SHA256SUMS. Start the matching runtime with shard 00001; it will read the other shards beside it. The runtime guide has the tested build and launch settings.

The Qwen Community License 1.0 is included as LICENSE. The runtime patch carries the llama.cpp project license in runtime/LICENSE.