122B-A10B 4bit
Hi,
Running Qwen3.6-27B MLX on M4 Max 128GB via LM Studio. Looking to move to Qwen3.5-122B-A10B — the MoE activation ratio makes it a better fit for this hardware than the dense 27B.
The existing MLX conversions either drop the vision tower entirely or hit the transformers ≥5.4.0 / Qwen2VLImageProcessor PyTorch dependency issue in pure MLX environments.
You already solved the equivalent problem on Mistral-Medium-3.5-128B (mlx-vlm mistral3 sanitize bug, vision weights preserved at full precision). Any chance you'd apply the same treatment to Qwen3.5-122B-A10B?
Happy to test and report back.
Yep, since this is just a Jinja template, it works independently of model quantization or size. As long as your inference engine supports Jinja chat templates, you are good to go! Going to close this out.