brayniac

/

Qwen3.5-35B-A3B-heretic

like 17

Hi,
Just wondering if there were any changes with the *.experts.gate_up_proj weight names during the decensoring that might cause a change in behaviour between vanilla Qwen 3.5 and the heretic'd version? I'm able to convert the normal weights to MLX without issue, and your 27B non-moe works perfectly (thanks very much for that, btw), but I'm not able to convert this model. Would appreciate any insights.
Thanks again,
Will