Qwen3.6 fully unlocked

A Qwen3.6-35B-A3B Aggressive variant has been released. While keeping the same MoE scale as the existing 3.5-35B release, it uses the newer Qwen 3.6 base.

The key point is the no refusals setting. The author states it recorded 0/465 refusals, and explains that it retains the original Qwen's capabilities as-is without changing the personality.

The files provided are as follows.

• Q8_K_P, Q6_K_P, Q5_K_P, Q4_K_P
• Q4_K_M, IQ4_NL, IQ4_XS
• Q3_K_P, IQ3_M, Q2_K_P, IQ2_M
• Includes mmproj, enabling vision support
• All quants were generated with imatrix

Specifications were also provided.

• 35B total / about 3B active
• 256 experts, 8 routed per token
• 262K context
• Multimodal: text + image + video
• hybrid attention: linear + softmax ratio 3:1
• 40 layers

Sampling parameters for testing were also shared.

• temp=1.0
• top_k=20
• repeat_penalty=1
• presence_penalty=1.5
• top_p=0.95
• min_p=0

Additionally, thinking mode can be turned off with enable_thinking=false, and llama.cpp requires the --jinja flag. In LM Studio, K_P quants may show as ? in the quant column, but it was noted that this is just a display issue and doesn't cause any problems with loading or running.