Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-PRISMAQUANT-PRISMASCOUT-8bit

Qwen3.6-27B-uncensored-heretic-v2 · Native MTP · PrismaSCOUT 8-bit

Mixed-precision quantization of llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved (an abliterated Qwen/Qwen3.6-27B), made with PrismaQuant using the PrismaSCOUT allocator at an 8-bit budget. It is the higher-fidelity, larger sibling of the 6-bit variant.

[!IMPORTANT]

• Runs only on vLLM through its stock compressed-tensors support (no GGUF, no llama.cpp, no custom kernels).
• Some layers are NVFP4, so this needs an NVIDIA Blackwell GPU.

Quantization

PrismaSCOUT generates many candidate per-layer format assignments, measures each one's end-to-end KL divergence from the original model on held-out text, and picks from the resulting Pareto frontier. That way format choices are based on how the whole model actually behaves, not on per-layer estimates. See the PrismaQuant repo for details.

Base model

• Architecture: Qwen3_5ForConditionalGeneration. It has 64 decoder layers (three linear-attention layers for every full-attention layer), a vision encoder and 1 MTP layer. The max position embeddings value is 262,144.
• Abliteration: Heretic v1.3.0 (an MPOA variant). The upstream card reports 6/100 refusals, down from 92/100 in the original. It also reports a KL divergence of 0.0021 and MMLU going from 86.65% to 85.67%. These figures were measured on the upstream BF16 model, not on this quant.
• Native MTP preserved: the Multi-Token Prediction weights were not changed by abliteration. They are kept unquantized here, so MTP speculative decoding works.

Usage (vLLM)

vllm serve IstroSec/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-PRISMAQUANT-PRISMASCOUT-8bit \
  --quantization compressed-tensors \
  --max-model-len 32768 \
  --kv-cache-dtype fp8 \
  --enable-prefix-caching \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

This is an example command. Change the context length and the number of speculative tokens to suit your hardware.

Other variants

• 6-bit variant of this model (27B)
• 35B-A3B MoE, 6-bit

Disclaimer

This model has had its refusal behaviour removed and can produce harmful or inappropriate content. You are responsible for how you use it and for adding any safeguards your use case needs.

Credits

Qwen (original model) · llmfan46 (Heretic abliteration, MTP preservation) · p-e-w/heretic · RobTand/prismaquant