Sign up

Wiself/Dark-Scarlett-v1.0-QAT-26B-A4B-Voice

🤗 Hugging Face sourcetext-generationapache-2.01.5 GBGGUF✓ HF checksums availableupdated today
No torrent yet →

Dark Scarlett's voice, matched to your QAT model — without training anything.

Dark-Scarlett-v1.0-QAT-26B-A4B-Voice

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

This carries ReadyArt/Dark-Scarlett-v1.0-26B-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 + Scarlett

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).

Scarlett changes how it writes. In our test, the Scarlett delta fused onto a QAT-Q4_0 head moved diction 96% with zero comprehension cost (single probe — see the Voice tool repo).

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-scarlett.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-scarlett.gguf --jinja

No QAT finetune was needed. No extra model was downloaded. Your model, darker 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 Scarlett's voice to the QAT head — correct math, matched foundation.

Your GGUF is built from Use
google/gemma-4-26B-A4B-it (standard) Scarlett voice
google/gemma-4-26B-A4B-it-qat-q4_0-unquantized this voice

The abliterated case

The most popular QAT-adjacent models are abliterated/uncensored variants. We tested this exact pairing class (Scarlett-delta → QAT-Q4_0 head):

  • Direct cast risks loops on heavily modified targets — repetitive output, unusable.
  • Delta cast — clean output, no loops, Scarlett phrasing intact, diction moved 96% with zero comprehension cost.

On abliterated targets, the delta path is the default (base is the normal instruct — correct math):

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: a delta carries only the difference from the base head, so the shift to MoE routing is minimal. Full story in the Voice tool card.

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

Target Works?
QAT-derived Gemma 4 26B A4B GGUFs (any quant) ✅ primary target
Abliterated QAT variants ✅ via the delta path above
Standard (non-QAT) 26B A4B GGUFs ✅ works, but the Scarlett voice is the exact match
Gemma 4 other sizes / non-Gemma ❌ shape mismatch

Notes

  • Sampler tips from the source card: Top-P 0.92, temp 1.0, no frequency/presence penalty.
  • Usage terms carry over: personal use only, 18+, you accept full responsibility for outputs — 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