Qwen3.5 35B A3B uncensored model drops with full MTP preservation
Uncensored Qwen3.5 with 785 MTPs retained launches in 5 formats on HuggingFace.
Get AI news that actually matters
One email a day. Zero fluff. Join 10,000+ professionals.
Deep Dive
LLMFan46 has released Qwen3.5-35B-A3B-uncensored-heretic-v2 with Native MTP Preserved, available in Safetensors, GGUF, NVFP4, and GPTQ-Int4 formats. The model shows a KL divergence of 0.0487 with an accuracy loss of 0.40%. It is intended for general-purpose AI assistance, while Qwen3.6 models are mainly for agentic and coding tasks. A benchmark is also included.
Key Points
- Retains all 785 MTPs (Multi-Token Prediction heads) for speculative decoding, available in Safetensors, GGUF, NVFP4, and GPTQ-Int4 formats.
- KL divergence of 0.0487 with only 0.40% accuracy loss, demonstrating low impact from abliteration compared to Qwen3.6's higher sensitivity.
- Optimized for general-purpose AI assistance, complementing Qwen3.6's focus on agentic and coding use cases.
Why It Matters
Provides a high-quality uncensored general-purpose model with minimal performance loss, ideal for local AI experimentation and deployment.