DarkIdol-Qwen3.8-27B-v1.1-Voice
One voice, three Qwen generations. Compatible with any Qwen 3.8 27B finetune or merge — plus we personally tested on 3.6 and 3.5 (via the delta), both with great results.
This is the voice of aifeifei798/DarkIdol-Qwen3.8-27B-v1.1 in a single portable tensor (~2.5 GB). If you already have a Qwen 27B GGUF — 3.5, 3.6, or 3.8 — cast it on with the Voice tool and hear how it reads: your base underneath, DarkIdol's writing on top.
Why this exists
DarkIdol, in the author's words, is built for "high-immersion roleplay" with a "Roleplay-First" philosophy — "Strict Anti-Puppeteering", "Organic Antagonists", "Literary Vernacular" (source card). That is the original model. We lift lm_head.weight from it — one tensor, ~2.5 GB — and cast it onto your base. Hear how it reads.
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 Qwen 27B GGUF you already have
voice cast ./qwen3.8-27b-Q4_K_M.gguf voice.safetensors --out ./voiced/qwen3.8-27b-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/qwen3.8-27b-darkidol.gguf --jinja
One file out. No adapter at runtime, no LoRA, no second model.
Works on any Qwen 3.8 27B — and two generations back
Compatible with any Qwen 3.8 27B finetune or merge (shape-gated; cast fails loud if vocab/hidden don't match). Then we kept going:
| 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, great results — standard voice |
| Heretic/abliterated Qwen 27B builds | ➡️ see delta variant — Q3.5-BlueStar-v2-27B-ultra-uncensored-heretic-v2.i1-IQ4_XS.gguf tested there |
| Other sizes / non-Qwen | ❌ shape mismatch — use the matching voice |
No model to try it on? Start with the Qwen 3.6 27B base — great results there. Loops on a heretic target? The delta variant has you covered.
The cast checks shape before touching anything and tells you plainly if it won't fit.
On heavily modified targets a direct cast can loop. The fix is the delta variant — it carries only the difference from the base instruct, so routing stays calm:
voice delta voice.safetensors --base Qwen/Qwen3.8-27B
voice cast ./model.gguf delta-voice.safetensors --out ./voiced/model.gguf
What's inside
voice.safetensors— thelm_head.weighttensor, BF16, arch-derived shape[248320, 5120], ~2.5 GBvoice.json— metadata: source, dtype, shape, architecture
Confirm after download: voice info voice.safetensors → lm_head.weight · [248320, 5120] · BF16.
Notes
- Sampler tips from the source card: temp 1.0, MinP 0.05, DRY sampler (0.8 / 1.75 / 2), medium reasoning effort, up to 64K context.
- Usage terms carry over: fiction-writing instrument, research use, sole end-user liability — read the source card's AUP before sharing voiced models. (The source frontmatter is apache-2.0; its body text additionally inherits Qwen research-license terms.)
- Casting quantizes only the head to Q8_0 (near-lossless); all other tensors byte-copied.
- Need a ready-made GGUF of the source? mradermacher's conversions, linked from the source card.
Credits
aifeifei798 — the DarkIdol finetune and its RP-first design. The Voice tool lifts and casts it. Qwen3.8 is Alibaba's model under its own terms.