Question about tokenization artifacts with some MLX models

[alexgavrisco]

Hey community,

what’s the proper place to report/discuss issues that seem to be specific to MLX-community models?

I’ve tried Devstral 2 Small from MLX-community and it produces some weird tokenization artifacts. At first I’ve been thinking that the issue is with my env/settings, but it seems that other folks encounter the same issue - lmstudio-ai/lmstudio-bug-tracker#1302

[tk1475]

Two parts:

Where to report: mlx-community models are conversions uploaded to the Hugging Face mlx-community org, so a problem with one specific converted model (Devstral 2 Small here) belongs on that model's HF page under its Community tab, not the mlx core repo. mlx is the array framework and doesn't own the conversions or tokenizers.

On the artifacts: tokenizer glitches with a converted model are almost always in the shipped tokenizer/chat-template or the mlx-lm version doing the decode, not mlx core. Worth checking:

• pip install -U mlx-lm — detokenization fixes land there regularly.
• Compare the model's tokenizer_config.json / chat_template against the original non-MLX repo; a bad conversion tends to show up as stray spaces, a doubled BOS, or special-token leakage.
• Run the original (non-MLX) model through mlx-lm to isolate conversion vs. your settings.

If it reproduces cleanly and looks like a decode bug rather than a bad upload, file it on ml-explore/mlx-lm with the model id and a minimal prompt. The LM Studio thread you linked points at the model/tokenizer, so I'd start on the mlx-community model page.