Add MiMo-V2.5-ASR STT support

[ailuntx]

Add MiMo-V2.5-ASR support to mlx-audio STT.

Changes:

• add a new mimo_v2_asr model family
• detect MiMo checkpoints in mlx_audio.stt.load()
• resolve the external MiMo-Audio-Tokenizer dependency from mlx_manifest.json
• support both raw HF tokenizer weights and the MLX tokenizer export layout
• add README and docs index entries

Validation:

• intention.wav -> Intention.
• conversational_a.wav -> expected coffee / Kaldi paragraph
• CLI JSON output path works for MiMo

Closes #718.

[Blaizzy]

Could you remove the "MLX" from the name?

The model repo already has tags to identify it as a supported model :)

[Blaizzy]

No need for a separate file for this.

[Blaizzy]

You can make this into sanitize method, like we do for every other model

[Blaizzy]

Please revert this

[Blaizzy]

Please revert all changes to this file except line 16

[Blaizzy]

The model will be auto discovered with the simple mapping

[Blaizzy]

Looks good overall just a few nits before we can merge

[ailuntx]

Addressed.

Changes in this update:

• removed the separate weight_loader.py
• reverted the custom mlx_audio/stt/utils.py loading path and kept a simple mimo remapping entry
• changed the displayed model name to MiMo-V2.5-ASR

I also added a small generic load_weights_from_path hook in base_load_model.
Without that hook, the default merged-weight path was spiking memory on the sharded MiMo checkpoint and hanging during local validation. The MiMo model now loads its main shards sequentially while still using the standard load() entry point.

Re-validated locally after the change:

• load(.../MiMo-V2.5-ASR-MLX)
• intention.wav -> Intention.
• CLI JSON output still works

[Blaizzy]

Thanks @ailuntx for the patience!

To make it easier, I will make the changes and you can look at what to improve in future PRs.

I would suggest asking the agent to follow the format of other existing models and not edit core files like the utils and loading paths.

[ailuntx]

Updated the branch to keep the MiMo changes within the model integration path.

Changes since your comment:

• removed the base_load_model / core loading-path hook
• moved MiMo key mapping into mimo_v2_asr.Model.sanitize
• added model_quant_predicate so the existing loader handles quantization
• ran pre-commit run --all-files successfully
• ran python -m compileall -q mlx_audio/stt/models/mimo_v2_asr mlx_audio/utils.py mlx_audio/stt/utils.py

The latest workflow runs are currently marked action_required, so they need approval before GitHub runs them.