Inaccurate Coordinate Outputs for MLX-Quantized UI-TARS-1.5 (4bit/6bit)

[francedot]

trafficstars

We've tested the quantized version of the UI-TARS-1.5 model (4-bit and 6-bit quantization) implemented with MLX. The work-in-progress implementation can be found here:
https://github.com/trycua/cua/tree/feature/agent/uitars-mlx

Problems Observed:

• The model struggles to output accurate (x, y) click coordinates when running under MLX quantization (both 4-bit and 6-bit).
• Specifically, the outputs often incorrectly target areas such as the center of the screen
• It’s possible that the quantized models are particularly sensitive to slight implementation differences compared to the full precision models (tested on cloud inference).

Testing Setup:

• Models tested locally on Mac with MLX (Apple Silicon).
• Observed significant performance hit when running model + c/ua VM simultaneously.
• Full precision model hosted on AWS endpoint produced correct behavior, suggesting the issue is specific to the MLX quantization.

Artifacts:

• Trajectories comparison archive (full precision vs 6-bit quantized outputs) attached:
  cua_uitars_trajectories.zip

Environment Details:

• Mac M3 Pro (local testing)
• AWS EC2 for cloud model hosting (full precision comparison)
• Cua framework for running agents (https://github.com/trycua/cua)

cua uitars trajectories.zip

[prncvrm]

@francedot this should be similar to this https://github.com/Blaizzy/mlx-vlm/pull/319

[prncvrm]

@francedot can you take a pull from my fork and validate if it works as expected?

[ddupont808]

@francedot can you take a pull from my fork and validate if it works as expected?

i pulled from your fork (commit b355cf1a85170957f626fa58c4c8dd205c92dd64 from today), but i am still noticing a bug with the coordinate outputs:

mlx-vlm ui-tars:

https://github.com/user-attachments/assets/a133b909-0385-46e9-854c-d351d2872813

torch ui-tars:

https://github.com/user-attachments/assets/4c7d6e5c-b167-4189-81bf-5b50492bdf27

misc environment details:

• using this code for running the agent: https://github.com/trycua/cua/commit/6a6fe48dbca0bd8f17652c538e08183ba289eefe
• and the following prompt: "please drag a line from the red circle to the green circle, then open a new tab and go to reddit\n\n(You are operating on macOS, use 'cmd' instead of 'ctrl' for most shortcuts e.g., hotkey(key='cmd c') for copy, hotkey(key='cmd v') for paste, hotkey(key='cmd t') for new tab).)"

[GaleiqTesting]

hey @prncvrm any fix for the above? keen on trying to get the mlx model to work

[prncvrm]

yes, the new UITars1.5 uses qwen2.5VL, while the fix i've raised was for Qwen2VL they are similar issue, i'll try to fix and raise a PR by this weekend hopefully

[Blaizzy]

Awesome, feel free to ping me when it ready!

I also left a comment on your current PR.

[prncvrm]

https://github.com/Blaizzy/mlx-vlm/pull/349 the PR's here, thanks to @ddupont808 🚀