mlasli/Qwen3.8-27B-Heretic-Uncensored-IQ2_M-GGUF

Qwen3.8-27B Heretic v2.1.0 (GGUF)

GGUF quantization of mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16 (v2.1.0). Full methodology, benchmarks, and the 3-way eval are documented there.

v2.1.0 fixes v2.0.0's false "uncensored" claim: the sweep now optimizes a 3-way direct / deflect / refuse judge (not a keyword refusal count) and the BF16 checkpoint scores 92% direct / 3% refuse / 5% deflect on the held-out harmful set (thinking OFF), with 1% benign over-refusal.

What's different about these GGUFs

• —MTP retained — the draft head (blk.64) is present and pinned to Q8_0 in every quant so speculative decoding stays accurate (block_count = 65).
• —imatrix-quantized — calibrated on wikitext-103, regenerated for the v2.1.0 weights.
• —Smoke-tested — every quant is loaded through llama-server --spec-type draft-mtp (MTP confirmed engaged) and scored 3-way on 50 harmful + 20 benign (thinking OFF). Results below.

Quantizations

Smoke test (llama.cpp + MTP, thinking OFF)

Usage

Load with llama.cpp (arch qwen35), enabling MTP speculative decoding and thinking-OFF:

llama-server -m qwen3.8-27b-heretic-uncensored-Q8_0.gguf \
  --spec-type draft-mtp -ngl 99 -c 4096

Set enable_thinking: false in the chat template (or the equivalent flag in your client). Ollama users can ollama create from the GGUF directly.

Vision caveat: text-only validated (see the BF16 card). MLX (v2.1.0): 6-bit and 8-bit Apple-Silicon quants are available. Thinking-OFF is required for the uncensored behavior — thinking-ON re-hedges/refuses more and can emit empty replies (documented in the BF16 card).

Abliteration removes safety alignment. Use responsibly and in accordance with your local laws and the upstream Apache-2.0 license.