Qwen3.6-35B-A3B-StyleTune-Voice
The StyleTune voice for Qwen's 35B MoE. One tensor. ~1.0 GB.
This is not a model. It is the voice of Gryphe/Qwen3.6-35B-A3B-StyleTune, extracted into a single portable tensor. Pair it with the Voice tool and any Qwen3.6-35B-A3B GGUF you already have — quants, abliterated variants, other finetunes — becomes StyleTune-voiced. Reasoning, knowledge, and MoE routing untouched.
Why this exists
A full 35B finetune is tens of gigabytes. The voice is ~1.0 GB — and it carries the entire style. The style lives in one tensor: the lm_head output projection, the last stop before text appears. Download the voice, not the model.
Two steps
# 1. Get the voice tool (one-time): https://huggingface.co/Wiself/voice
python3 voice.py path # optional, puts `voice` on your PATH
# 2. Cast it onto any Qwen3.6-35B-A3B GGUF you already have
voice cast ./qwen3.6-35b-a3b-Q4_K_M.gguf voice.safetensors --out ./voiced/qwen3.6-35b-a3b-styletune.gguf
Append --speak to force the voice through output.weight (invented if the target lacks it and geometry survives; GGUF only).
Run it:
llama serve -m ./voiced/qwen3.6-35b-a3b-styletune.gguf --jinja
That's it. One file out, nothing extra at runtime. No adapter, no LoRA, no second model.
What's inside
voice.safetensors— thelm_head.weighttensor, BF16, shape[248320, 2048], ~1.0 GBvoice.json— metadata: source, dtype, shape, architecture
Bit-for-bit identical to the tensor that makes Gryphe/Qwen3.6-35B-A3B-StyleTune sound the way it does. If you have the original finetune, you don't need this file. If you have any other Qwen3.6-35B-A3B — a quant, an abliterated variant, a different finetune — this voice drops in.
Compatibility
| Target | Works? |
|---|---|
| Any Qwen3.6-35B-A3B GGUF (any quant: Q4_K_M, Q5_K_M, Q8_0…) | ✅ |
| Abliterated / uncensored 35B-A3B GGUFs | ✅ (use the delta path below if it loops) |
| Other Qwen3 sizes / Qwen3.5 and below | ❌ vocab/hidden mismatch — use the matching voice |
| Non-Qwen architectures | ❌ |
The cast checks shape before touching anything and tells you plainly if it won't fit.
On some abliterated targets a direct cast can loop. The fix is the delta variant — it carries only the difference from the base instruct, so MoE routing stays calm:
voice delta voice.safetensors # needs the base: --base Qwen/Qwen3.6-35B-A3B
voice cast ./model.gguf delta-voice.safetensors --out ./voiced/model.gguf
This is the exact path that fixed looping on abliterated QAT models — see the Voice tool card for that story.
Notes
- Quality: the voice is stored at the original BF16 precision. Casting to GGUF quantizes only the head to Q8_0 (≈¼ size, near-lossless); every other tensor is byte-copied from your model.
- Casting to safetensors targets works too — the head is replaced at the target's own dtype, no quantization.
- MoE note: Qwen3.6-35B-A3B routes across experts per token. The voice only replaces the output projection after routing, so expert behavior is preserved — style changes, reasoning doesn't.
- Verify after download:
voice info voice.safetensors→ should readlm_head.weight · [248320, 2048] · BF16.
Credits
Gryphe trained the StyleTune and proved one tensor changes everything. The Voice tool lifts and casts it. Qwen3.6 is Alibaba's model under its own license — this voice carries the source model's license lineage; check the source card before sharing voiced models.