AI BriefingQwen3.6 35B Heretic released
Key point
A decensored version of Qwen3.6 35B A3B using Heretic has been released.
Details
A decensored version of Qwen/Qwen3.6-35B-A3B, created as an MPOA variant of Heretic v1.2.0, has been released.
As key metrics, it claims to have reduced the refusal response rate from 83/100 → 10/100 while maintaining a KL divergence of 0.0015.
The abliteration targets presented are the following 3:
- attn.o_proj
- attn.out_proj
- mlp.down_proj
Additionally, MMLU test results and a detailed parameter table are included, but the focus of the post is on distributing a decensored version with reduced refusals rather than performance comparison.
However, since the post strongly mixes in model promotion and sponsorship requests, for news value it seems appropriate to selectively extract only the new model variant and numerical comparisons.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images