MLX does not run Gemma 4 unified architecture

[freerainboxbox]

Which version of LM Studio?
0.4.16+1 (MLX runtime 1.8.5)

Which operating system?
macOS Sequoia 15.6.1 (M1 Max 64 GB)

What is the bug?
MLX quants of Gemma 4 12B, or any potential model using the unified architecture do not load with the MLX runtime.

Screenshots
N/A

Logs

Error when loading model: ValueError: Model type gemma4_unified not supported.
Error: No module named 'mlx_vlm.speculative.drafters.gemma4_unified'

To Reproduce
Steps to reproduce the behavior:

1. Download https://huggingface.co/mlx-community/gemma-4-12B-it-qat-4bit OR quantize directly from source using python -m mlx_vlm convert --hf-path google/gemma-4-12B-it-qat-q4_0-unquantized -q --q-group-size 32 --q-bits 4
2. Attempt to load the new model

[kanishk619]

Same issue on beta 0.4.16+1 with all runtimes up to date on beta channel as well

[boyhood-fallers0v]

Error: No module named 'mlx_vlm.speculative.drafters.gemma4_unified'

[ijedgington-blip]

why dont they update mxl-lm and mlx-vlm with there releases, i can load them from the command line but not in lm studio

[rohitst]

same issue

[bejminliou]

same.

mlx-community/gemma-4-12B-it-4bit

HTTP 500 {
  "detail": "VLM load failed: Model type gemma4_unified not supported. 
             Error: No module named 'mlx_vlm.speculative.drafters.gemma4_unified';
             LLM fallback also failed: Model type gemma4_unified not supported."
}

[MarianSWA]

Same here! Couldn't this be automated somehow so we don't need to wait for LM Studio to manually update the versions every single time there's an update?

[plz12345]

It could be automated, but they probably view it as risky. This isn't just a new model, it's a new architecture, so it has not been tested by the LMS team within the app. Allowing blind architecture updates is a recipe for failure.

Now, that said, why LMS doesn't block the very easy download of incompatible models is another story.

[Playitcooool]

Same even for their own models:lmstudio-community:gemma-4-12B-it-MLX-4bit.

It seems all models that tagged as "gemma4_unified" cannot be loaded correctly.

[bejminliou]

The fix is already merged upstream in mlx-vlm and will hopefully be available in the next update once they bump their dependency.

I agree with @plz12345 on this one. Auto updating dependencies risks breaking the application (internal APIs, function signatures, expected data structures, ETC. all subject to change). So they must manually test and rebuild before releasing updates. It could be a week, maybe a month from now. Stay strong bros.

[jlrouzies-fr]

Same here, with MLX 1.9.1

026-06-26 10:19:44 [DEBUG] [LLMProcess] Failed to load model _0x2ff51d [Error]: Failed to load model. at _0xc273ce.loadModel (/Applications/LM Studio.app/Contents/Resources/app/.webpack/lib/llmworker.js:1:562603) at process.processTicksAndRejections (node:internal/process/task_queues:104:5) at async _0xc273ce.handleMessage (/Applications/LM Studio.app/Contents/Resources/app/.webpack/lib/llmworker.js:1:554557) { cause: 'Error when loading model: ValueError: Model type gemma4_unified not supported.', suggestion: undefined, errorData: undefined, data: undefined, displayData: undefined, title: 'Failed to load model.' }