Wide performance discrepancy between LM Studio MLX and `mlx_vlm.generate` with Qwen3.5-35B-A3B

[Belluxx]

When running Qwen3.5-35B-A3B 4bit MLX on an M4, I noticed that using a ~3000 tokens prompt:

• LM Studio gets consistently around 78 tok/s token generation and 6s TTFT
• mlx_vlm.generate reaches 108 tok/s token generation and 2s TTFT

The discrepancy in token generation is almost identical even with very short prompts (<100 tokens).

[hlibr]

I've also noticed that TTFT seems to increase with subsequent turns (which is not the case with GGUFs, but I don't know if this is mlx-vlm behavior or LM Studio specific).