llmfan46's Qwen3.6 27B uncensored model retains full 15 MTP, 6% refusal rate
New uncensored 27B model with native multi-token prediction and minimal refusals
Deep Dive
llmfan46 released Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved, a 27B parameter fine-tune of Qwen3.6. All variants are confirmed to have their full 15 MTPs retained and preserved. Comes with benchmark too. Available in GGUF and NVFP4 formats.
Key Points
- 27B model with 94% uncensored response rate (only 6 refusals per 100 prompts).
- Retains all 15 native multi-token prediction (MTP) heads with a KLD of 0.0021 from base model.
- Available in Safetensors, GGUF, and NVFP4 formats, all with full MTP preservation.
Why It Matters
Uncensored 27B model with preserved MTP enables high-quality local AI without safety filters.