Boulesis-26B-A4B-QAT-Voice

Boulesis's voice, fitted to QAT-derived GGUFs. Same idea as the main voice, correct math for QAT heads.

This carries SubMaroon/Boulesis-26B-A4B as a delta against the standard instruct, matched to Google's QAT checkpoint (google/gemma-4-26B-A4B-it-qat-q4_0-unquantized). If your GGUF is built from QAT weights, use this one; otherwise the standard voice is the exact fit.

QAT + Boulesis

Two separate things, stacked:

Google's QAT trains the model to survive quantization — 54% less perplexity drop at Q4_0 versus standard post-training quantization (Google Developers Blog).

Boulesis changes how it writes. The author's own build log attributes the biggest roleplay gain to a single transplanted tensor — the lm_head from StyleTune-V2 — and this delta carries exactly that shift, measured against the standard instruct head. (Our own probe numbers don't exist yet; the linked CaliperBench result — best RP in the 24–26B class, thinking mode, September 2026 — is the source author's, not ours.)

If you already run a QAT GGUF, this variant keeps the QAT foundation intact instead of overwriting it with standard-instruct weights.

Two steps

# 1. Get the voice tool (one-time): https://huggingface.co/Wiself/voice
python3 voice.py path

# 2. Cast onto your QAT-derived GGUF
voice cast ./gemma-4-26b-a4b-it-qat-q4_0-Q4_0.gguf voice.safetensors --out ./voiced/gemma-4-26b-a4b-qat-boulesis.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-qat-boulesis.gguf --jinja

No QAT finetune was needed. No extra model was downloaded. Your model, a more decisive voice, less quantization loss.

Why the QAT-matched variant exists

The QAT checkpoint's lm_head weights differ slightly from the standard instruct's (that's the point of QAT — the weights learned to live with quantization). This voice is cast as a delta against the normal google/gemma-4-26B-A4B-it (voice − base where base is the standard instruct, not QAT). When you later do delta + QAT_head → Q8_0, you add Boulesis's voice to the QAT head — correct math, matched foundation.

The abliterated case

The most popular QAT-adjacent models are abliterated/uncensored variants. On those, prefer the delta path — a delta carries only the difference from the base head, so the shift to MoE routing is minimal and heavily modified targets stay calm:

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

Why it works: full story in the Voice tool card. Note the base here is tied (no separate head tensor) — needs a current Voice tool, which deltas the head against the tied embedding automatically.

What's inside

• voice.safetensors — the voice delta (voice − base), source dtype preserved, arch-derived shape [262144, 2816], ~1.5 GB
• voice.json — metadata: source, dtype, shape, base

Confirm after download: voice info voice.safetensors → [262144, 2816] (delta marker present).

Compatibility

Notes

• Sampler tips from the source card: thinking ON, temp 1.0, top-K 64, top-P 0.95, repetition penalty 1.05–1.1.
• Usage terms carry over: the source is Gemma-licensed — see the source card before sharing voiced models.
• Casting quantizes only the head to Q8_0 (near-lossless); every other tensor is byte-copied — your QAT weights stay QAT weights.

References & Credits

• Voice source: SubMaroon/Boulesis-26B-A4B — QK task arithmetic + fused LoRA composite (body: Heretic; head: StyleTune-V2; reasoning graft: Pantheon-1.1).
• QAT: Google Developers Blog — checkpoint google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, Gemma terms.
• Tool: Voice — lift a voice, cast it onto any compatible GGUF.