A.I.
More labs
Mistral Large 4
Mistral Large 4 is the flagship model of Mistral AI, a French lab. It went into public preview on 6 October 2026, and Mistral says the weights, the downloadable model files, will be released by the end of that month (Mistral, 6 Oct 2026).
- Status
- Preview
- Price
- $1.36 in, $4.18 out (list)US dollars per million tokens (word pieces) in and out, unless noted
- Released
- 6 Oct 2026
- Last checked
- 12 Oct 2026newest source date on this page
Key facts
| Field | Mistral Large 4 |
|---|---|
| Size | 1.05 trillion parameters, 52 billion active, plus a 1.6 billion parameter vision encoder |
| Design | Mixture of Experts |
| Context | 1M tokens |
| Inputs | Text and images |
| Price, list | $1.36 in, $4.18 out, $0.14 cached input |
| Price, preview sale | $0.68 in, $2.09 out, $0.07 cached input |
| Weights | Promised by end of October 2026 |
Size and price are from Mistral's model docs (Mistral docs); the announcement rounds the size to 1 trillion. A parameter is one learned number in the model. Mixture of Experts means only part of the model, the 52 billion active parameters, runs for each token. A token is a chunk of text, often part of a word. Prices are US dollars per million tokens. Cached input is repeated text the service has already read, billed cheaper.
Training
Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, with training data in more than 160 languages.
What Mistral claims
- 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, all coding tests.
- Top five on the Artificial Analysis Cyber Index, which tests finding and fixing security flaws.
- These are Mistral's own reported scores. Mistral's announcement says several closed models score near zero on one cyber test because they refuse the task.