Mistral Large 4 is an API today and open weights by month's end. Use the gap to decide who runs it

Most enterprise model decisions force an early commitment. Rent a hosted API and accept the provider’s custody of your prompts, or download open weights and accept the work of running them. Mistral Large 4, released in public preview on October 6, briefly offers both, one after the other: an API now, and downloadable weights that Mistral says will arrive by the end of the month.

That sequence is useful, if a team treats the weeks in between as an evaluation window rather than a decision already made.

What Mistral released

Large 4 is a mixture-of-experts model with about 1 trillion total parameters and 52 billion active per token. It accepts text and images. Mistral prices the API preview at $1.36 per million input tokens and $4.18 per million output tokens, and points the model at cybersecurity, software engineering, agentic workflows, finance, law, manufacturing and scientific work.

Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, and serves it in several regions, including a European deployment that Mistral operates end to end, “independently of other digital service providers and under European law”. The benchmark results in the announcement, such as 82% on a vulnerability patching test and 93% on Cybench, are Mistral’s own.

Two caveats are written into the announcement itself. The weights are coming, not here, and Mistral has not yet published their license. And reinforcement learning training is “still in flight”.

Four ways to run the same model, by who holds custody

Mistral operatesYou operate

1. Mistral API, any regionAvailable in preview
2. Mistral's European deploymentMistral operated, under European law
3. Weights in your private cloudAfter release, subject to license
4. Weights in your data centerFull custody, full operating duty

The model you test is still changing

“Still in flight” has a practical meaning. The model a team evaluates through the API in October may not be the one whose weights it downloads at the end of the month, and it may not behave identically once served on a different stack with different quantization.

So the preview results are provisional. Freeze the evaluation set now, run it against the API, and run the same set again on the released weights, in the serving setup you would actually use. As we wrote about model retirement, a model swap is a release, and this one has a swap built into its launch.

Four questions to answer before the weights arrive

What does the license allow? Until it is published, no self-hosting plan is real. Commercial use, redistribution, fine-tuning and any usage restrictions decide whether the right-hand side of the line above exists for you at all.

Can you serve it? Only 52 billion parameters are active per token, but a standard deployment still keeps all trillion in memory, on the order of a terabyte of accelerator memory at 8-bit precision. That is a multi-GPU cluster, not a workstation.

Who operates it? This week, writing about Reflection’s Beam, we argued that with open weights, the operating duty moves to you: capacity, patching, monitoring and evaluation become your team’s job. Large 4 is a larger test of the same argument.

What price does it have to beat? The hosted preview sets a benchmark. A self-hosted deployment has to beat $1.36 and $4.18 per million tokens at your real volume, including idle capacity and the people who run it, or offer custody the API cannot.

The open weights are what make Large 4 newsworthy. For an enterprise, the more valuable part may be the month before them: a chance to measure the model on its own tasks while the hosted option is still the cheap one to try, and to decide who should run it with evidence in hand.