Open Source

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.

📬 Get the top 10 AI stories daily