The StyleTune V2 voice for Gemma 4 26B MoE — as a download.

Gemma-4-26B-A4B-StyleTune-V2-Voice

52% fewer clichés. 19.9% shared vocabulary. One tensor out of 659. ~1.5 GB.

This is not a model. It is the voice of Gryphe/Gemma-4-26B-A4B-StyleTune-V2, extracted into a single portable tensor. Pair it with the Voice tool and any Gemma 4 26B A4B GGUF you already have becomes StyleTune-voiced — a 52 GB download becomes a 1.5 GB download.

---

Why this exists

StyleTune V2 is a single-epoch finetune of exactly one tensor: the lm_head output projection. Gryphe found a second epoch does nasty things to MoE models, so V2 is the stable one — the metrics barely moved and stability improved dramatically. Everything that makes it sound like StyleTune lives in that one tensor, so that's all we lift.

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 26B A4B GGUF you already have
voice cast ./gemma-4-26b-a4b-it-Q4_K_M.gguf voice.safetensors --out ./voiced/gemma-4-26b-a4b-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-4-26b-a4b-styletune.gguf --jinja

One file out. No adapter at runtime, no LoRA, no second model.

---

The numbers

From Gryphe's benchmarks (source card) — 200 diverse roleplay prompts, greedy 0.0, versus the base instruct:

Metric	Base	StyleTune V2
Clichés per 100 words	1.141	0.551 (−52%)
Shared trigram vocabulary	—	19.9%

Our own independent spot-check (10 roleplay prompts, heuristic, not comparable — direction only) put shared trigram vocabulary at 1.6% (temp 1.0) / 3.2% (greedy) when cast onto a different 26B base — the voice transfers and stays distinct. Details in the Voice tool card.

---

Works on more than the base model

Because the voice is just a tensor, it casts onto any compatible Gemma 4 26B A4B GGUF:

• Different finetunes — keep the stronger base, gain the StyleTune voice
• Abliterated / uncensored variants — see loops below
• Quantized GGUFs — any quant; your tensors stay untouched, only the head is replaced

Proven on Orion

We cast this voice onto TheDrummer/Orion-26B-A4B-v1.4 — phrasing came out 98% different from the base (1.6% shared trigrams at temp 1.0, 3.2% at greedy), with no loops and no empty responses across 10 prompts. The voice transfers across finetune boundaries within the same architecture.

Loops on abliterated targets?

On some abliterated targets (we hit this on the QAT uncensored variant), a direct cast can loop. The delta variant fixes it — it carries only the difference from the base instruct, so MoE routing stays calm:

voice delta voice.safetensors --base google/gemma-4-26B-A4B-it
voice cast ./model.gguf delta-voice.safetensors --out ./voiced/model.gguf

That exact path turned the looping QAT-uncensored cast into a clean one. See the Voice tool card for the full story.

---

What's inside

• voice.safetensors — the lm_head.weight tensor, F16, shape [262144, 2816], ~1.5 GB
• voice.json — metadata: source, dtype, shape, architecture

Values match the tensor in Gryphe/Gemma-4-26B-A4B-StyleTune-V2 within F16 rounding — this artifact predates voice.py's source-dtype preservation, so it is F16 where the source is BF16. Current extractions keep the source dtype.

---

Compatibility

Target	Works?
Any Gemma 4 26B A4B GGUF (any quant)	✅
Gemma 4 other sizes (9B, 12B, 31B)	❌ shape mismatch — use the matching voice
Non-Gemma architectures	❌

---

Notes

• Precision: stored F16 (see above — predates dtype preservation). Casting quantizes only the head to Q8_0 (near-lossless); all other tensors byte-copied.
• QAT users: if your GGUF is built from Google's QAT checkpoint, use the QAT voice variant — same style, QAT-matched weights.
• Sampler tips from Gryphe: temp 1.0, MinP 0.10, DRY sampler on.
• Verify: voice info voice.safetensors → lm_head.weight · [262144, 2816] · F16.

---

Credits

Gryphe — StyleTune technique, benchmarks, and the discovery that one tensor carries the voice. Anthracite and Latitude — collaboration and data. The Voice tool lifts and casts it. Base model and QAT checkpoints are Google's, under Gemma terms — check the source card before sharing voiced models.