Gemma-4-12B-StyleTune-Voice
56% fewer clichés. 16.8% shared vocabulary. One tensor. ~2.0 GB.
This is not a model. It is the voice of Gryphe/Gemma-4-12B-StyleTune, extracted into a single portable tensor. Pair it with the Voice tool and any Gemma 4 12B GGUF you already have becomes StyleTune-voiced — reasoning, knowledge, and instruction following untouched.
Why this exists
A full StyleTune finetune is 24 GB. The voice is 2.0 GB — and it carries the entire style. The style lives in one tensor: the lm_head output projection, the last stop before text appears. Gryphe trained only that tensor; we lift only that tensor.
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 Gemma 4 12B GGUF you already have
voice cast ./gemma-4-12b-it-Q4_K_M.gguf voice.safetensors --out ./voiced/gemma-4-12b-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/gemma-4-12b-styletune.gguf --jinja
That's it. One file out, nothing extra at runtime. No adapter, no LoRA, no second model.
The numbers
From Gryphe's own benchmarks (source card) — 200 diverse roleplay prompts, greedy 0.0, versus the base instruct:
| Metric | Base | StyleTune |
|---|---|---|
| Clichés per 100 words | 1.050 | 0.463 (−56%) |
| Shared trigram vocabulary | — | 16.8% (an almost entirely different phrasing) |
Everything that makes StyleTune StyleTune is in this file. What does not change: reasoning, world knowledge, instruction following, multilingual capability — none of it lives in lm_head.
What's inside
voice.safetensors— thelm_head.weighttensor, BF16, shape[262144, 3840], ~2.0 GBvoice.json— metadata: source, dtype, shape, architecture
Bit-for-bit identical to the tensor that makes Gryphe/Gemma-4-12B-StyleTune sound the way it does. If you have the original model, you don't need this file. If you have any other Gemma 4 12B — a quant, an abliterated variant, a different finetune — this voice drops in.
Compatibility
| Target | Works? |
|---|---|
| Any Gemma 4 12B GGUF (any quant: Q4_K_M, Q5_K_M, Q8_0…) | ✅ |
| Abliterated / uncensored Gemma 4 12B GGUFs | ✅ (use the delta variant if it loops) |
| Gemma 4 other sizes (9B, 26B, 31B) | ❌ vocab/hidden mismatch — use the matching voice |
| Non-Gemma 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 google/gemma-4-12b-it
voice cast ./model.gguf delta-voice.safetensors --out ./voiced/model.gguf
This is the exact path that fixed looping on the 26B QAT abliterated model — 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.
- Sampler tips from Gryphe: temp 1.0, MinP 0.10, DRY sampler on. Gemma 4's native chat template applies automatically.
- Verify after download:
voice info voice.safetensors→ should readlm_head.weight · [262144, 3840] · BF16.
Credits
Gryphe trained the StyleTune and proved one tensor changes everything. The Voice tool lifts and casts it. Gemma 4 is Google's model, under its own terms — this voice carries the source model's license lineage; check the source card before sharing voiced models.