Support for Mistral-Large-3-675B-Instruct-2512 please

[bibproj2]

Two weeks ago Mistral AI released the Mistral 3 collection, which includes Ministral 3 (which you support already) and Mistral Large 3 (which is not yet supported).

Like Mistral 3 Small it has vision capability, which is why it probably fits in with mlx-vlm rather than mlx-lm.

---

Reason: Mistral is really good with translations to European languages. And some Asian langauges.

• The small 24B Mistral Small models generated some of the best translations to French. Which is probably not surprising as Mistral is a French company.
• The new Mistral Large 3 seems to produce the best Finnish translations.

I work on a large translation project. MLX support for this model would be a huge help.

---

Architecture: Like Kimi K2, Mistral Large 3 uses the "standard" DeepSeek V3-0324 architecture, with some minor variations.

This is confirmed by

• Sebastian Raschka drew the architecture diagrams comparing the DeepSeek V3/R1 model to the new Mistral Large 3 model.
• A comment by a Mistral employee compares the differences.

There are some (minor) differences:

• The model is not yet transformers-compatible, as confirmed by the Mistral employee. As such there is no config.json file.
• You might glean more information from the llama.cpp pull request 17730. It did not take them very long to get this working for llama.cpp.
• It does work with vLLM

They also have a 12B Eagle Speculator model for for speculative decoding. It uses NVIDIA’s Eagle3 speculative decoding approach with the Model Optimizer to predict a single draft token efficiently, making it useful for high-concurrency inference scenarios where fast token generation is a priority. I do not know if this could work with MLX.

[Blaizzy]

Hey @bibproj2

Thanks, that makes a lot of sense!

I have added it to the backlog 👌🏽

[bibproj2]

For info: I tried to get it working by myself in mlx-lm as I know this a bit better. But alas, not good enough yet.