Mistral Large 4: one trillion parameters, pricing and open weights
Mistral Large 4 has one trillion parameters and lights up only 52 billion. Price per million tokens, what it does and when the open weights arrive.
Mistral Large 4 is the new model from Mistral AI: it has one trillion parameters but lights up only 52 billion at a time. It was introduced on October 6, 2026, it is an API preview today, and the company says the open weights arrive at the end of the month so you can host it yourself.
One trillion parameters, but only 52 billion active
What matters in a model is not how much it weighs, but how much it lights up. Mistral Large 4 totals one trillion parameters, but it activates only 52 billion to answer. It is a model that does not use all its pieces at once: in practice, total size and the cost of each answer stop moving together.
What Mistral Large 4 does
- It is multimodal: it reads documents, charts and images.
- It speaks many languages: it was trained on more than 160 languages.
- It performs in cybersecurity: Mistral reports 93% on Cybench.
Heads up on that last number: the 93% is the result Mistral reports, not an independent measurement.
Mistral Large 4 API pricing
On its API preview it costs $1.36 per million input tokens and $4.18 per million output tokens. If you pay for another model today, compare against the real size of your inputs and outputs before deciding.
Open weights: what is still missing
The most interesting part is the promise to release the open model at the end of the month. Instead of renting an API, you could host it yourself, on your own infrastructure. But three things are pending today:
- It is an API preview, not a general release.
- The weights are not available yet.
- The license is not specified.
Without the license you will not know whether you can use it commercially, so do not plan anything serious until you see it.
How to apply it in your business?
This is a WAiBOT suggestion, not a Mistral feature: before switching models, run a simple test.
- Write down which model you use today and what you pay per million tokens.
- Take 20 to 30 real cases: documents, charts, images or text in several languages.
- Run them on the Mistral Large 4 API preview and compare quality and cost.
- If you decide to host it yourself, wait for the weights and check the license before committing infrastructure.
If you want to see how another model's cost compares, read our breakdown of Claude Haiku 5.5: pricing and API fine print.
Frequently asked questions
What is Mistral Large 4? It is the Mistral AI model introduced on October 6, 2026, with one trillion parameters and 52 billion active.
How much does Mistral Large 4 cost? On its API preview, $1.36 per million input tokens and $4.18 per million output tokens.
Is Mistral Large 4 open source? Not yet. Mistral says it will release the open weights at the end of the month, but the license is not specified.
How good is it on Cybench? Mistral reports 93%, a result that comes from the company itself.
Conclusion
Mistral Large 4 stands out for its balance between total size and active model, its pricing and the promise of open weights. Today it is an API preview, with results reported by Mistral and no license in sight. Tell us which model you use today and we will tell you if this one fits. Source: Mistral AI, Introducing Mistral Large 4.