Mistral Large 4 is open weight: what you can run, and what it costs on the API

What happened

Mistral published the weights of Mistral Large 4 on October 29, 2026. The model went into public preview on the API on October 6, with a promise of weights by the end of the month, and Mistral says three more weeks of reinforcement learning went in before this release.

The model card in Mistral's docs lists 52B active and 1.05T total parameters, a 1.6B vision encoder and a 1M-token context. It reads text and images and answers in text.

Can you run it yourself?

Not at full precision on any machine you'd buy for a home office. The main checkpoint, Mistral-Large-4-1T-A52B, takes about 1.06 TB on Hugging Face. A smaller NVFP4 variant for recent NVIDIA cards still takes about 610 GB. Hugging Face's launch post lists day-0 support in Transformers, vLLM and llama.cpp, plus GGUF quants from Unsloth and the llama.cpp team. Its Transformers example splits the model across GPUs with tensor and expert parallelism. At those sizes that's data-center hardware; the GGUF quants are the route to try on one large-memory workstation.

The license is the Mistral Ecosystem Open Weight License, which Hugging Face describes as an MIT derivative with limits on model-as-a-service offerings.

What it costs on Mistral's API

Those are the USD prices on the model card. For scale, Anthropic lists Opus 5.5 at $4 input and $20 output per million tokens. On output-heavy agent work that's roughly a tenth of the bill.

How good is it at code?

Close to the best open models, and behind Opus 5.5 in the one blind comparison. The numbers come from Mistral's release post.

Reasoning can't be switched off

The lowest setting is low: Mistral's three reasoning modes are low, medium and high, with high the default, and low replaces the old none. Hugging Face notes that passing none still runs at high. A pipeline that sent none to save tokens should switch to low, or it pays for full reasoning on every call.

What it means for you

1. API builders on a budget: run a batch of your real requests at low and medium before you move traffic.
2. Teams that must keep code in-house: put it on the shortlist next to GLM-5.3, if you have the GPUs and the license fits.

Mistral's own app has its Mistral Vibe plans and prices on a separate page.

Try it

I'll paste one task and three answers to it, made at reasoning low, medium and high.
1) Grade each answer pass / fail against the criteria below, with one line on why.
2) Name the lowest mode that passes.
3) If low fails, say exactly what it got wrong, so I can tell whether a better prompt would fix it.
Don't reward length. A short correct answer passes.

Task: [paste the task]
Pass criteria: [what a correct answer must contain]
Low: [paste]
Medium: [paste]
High: [paste]

Here is a bug report and the code it points at. Work in this order and don't skip a step:
1) Write the smallest test that fails because of this bug. Explain in one sentence why it fails.
2) Propose the smallest patch that makes the test pass.
3) List what else the patch could break, and one more test for each.
If you can't make the test fail, stop and tell me what information is missing. Don't patch a bug you couldn't reproduce.

Language and test framework: [e.g. TypeScript, Vitest]
Bug report: [paste]
Code: [paste the file or function]

Sources

• Mistral, Mistral Large 4 (Le Chonk) is now open weight, October 29, 2026
• Mistral, Introducing Mistral Large 4, October 6, 2026
• Mistral Docs, Mistral Large 4 model card, checked October 29, 2026
• Hugging Face, Welcome Le Chonk: Mistral Large 4 on Hugging Face, October 29, 2026