DarkIdol-Qwen3.8-27B-v1.1-Heretic-DELTA-Voice
Try the standard voice first — it's the more direct one. This delta is only for you if your target is heretic/abliterated and the direct cast loops.
This carries aifeifei798/DarkIdol-Qwen3.8-27B-v1.1 as a delta against the stock instruct (Qwen/Qwen3.8-27B): voice − base, matched so it fuses onto modified foundations instead of overwriting them.
Why the delta variant exists
A direct cast replaces the whole output head — on heavily modified (heretic/uncensored) targets that can shock routing into loops. A delta carries only the difference from the base head, so the shift is minimal and modified targets stay calm. Same voice, gentler landing.
The subtraction behind it is voice lm_head minus base lm_head — same tensor, same role — computed with a current Voice tool.
| Your situation | Use |
|---|---|
| Stock or lightly-modified Qwen 27B GGUF | Standard voice — direct, one step |
| Heretic/abliterated target, or direct cast loops | this delta — the path below |
Two steps
# 1. Get the voice tool (one-time): https://huggingface.co/Wiself/voice
python3 voice.py path
voice cast ./my-heretic-qwen3.8-27b-Q4_K_M.gguf voice.safetensors --out ./voiced/heretic-darkidol.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/heretic-darkidol.gguf --jinja
No finetune was needed. No extra model was downloaded. Your heretic base, DarkIdol's voice, no loops.
Same cross-generation compatibility
Same head geometry as the standard voice — we personally tested the cross-gen story here. Details on the standard card.
| Target | Result |
|---|---|
| Qwen 3.8 27B finetunes and merges (any quant) | ✅ compatible (shape-gated) |
Qwen 3.6 27B (Qwen3.6-27B-IQ4_NL.gguf, stock base) | ✅ tested with standard voice, great results |
Heretic Q3.5-BlueStar-v2-27B-ultra-uncensored-heretic-v2.i1-IQ4_XS.gguf | ✅ tested with this delta, great results |
What's inside
voice.safetensors— the voice delta (voice − base), source dtype preserved, arch-derived shape[248320, 5120], ~2.5 GBvoice.json— metadata: source, dtype, shape, base
Confirm after download: voice info voice.safetensors → [248320, 5120] (delta marker present).
Notes
- Sampler tips from the source card: temp 1.0, MinP 0.05, DRY sampler (0.8 / 1.75 / 2), medium reasoning effort.
- Usage terms carry over: fiction-writing instrument, research use, sole end-user liability — 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 base weights stay your base weights.
References & Credits
- Voice source: aifeifei798/DarkIdol-Qwen3.8-27B-v1.1 — RP-first DarkIdol finetune on Qwen3.8-27B.
- Base:
Qwen/Qwen3.8-27B— stock instruct, the subtraction baseline. - Tool: Voice — lift a voice, cast it onto any compatible GGUF.