lmstudio-ai / lmstudio-ai/mlx-engine

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

Open
#285 1 comment 0 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
1.2k
Forks
133
Avg merge
21h 6m
Merged PRs (30d)
1

Description

When running `Qwen3.5-35B-A3B` [4bit MLX](https://huggingface.co/mlx-community/Qwen3.5-35B-A3B-4bit) 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).

Contributor guide

Open the contributing guide

Research direction

Start by reproducing the Qwen3.5-35B-A3B 4bit benchmark on an M4 with the reported ~3000-token and short prompts, comparing LM Studio against mlx_vlm.generate. Investigate the generation-speed and time-to-first-token difference; done means identifying and documenting the cause or producing a focused fix with measurements showing the discrepancy is resolved.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.