Mistral unveils Large 4, an open 1T model

Mistral has released the public preview of Mistral Large 4, a multimodal model with one trillion parameters, 49 billion of them active in each response. The company describes it as its largest and most capable system to date, and plans to publish its weights at the end of October 2026.

You can already try it through the Mistral Studio API. The model combines text generation, reasoning, coding, tool use and image understanding in a single system.

What Mistral Large 4 offers

Mistral Large 4 uses a MoE architecture, short for mixture of experts. Instead of activating all its parameters for every task, it selects only the groups it needs. That makes it possible to work with a huge model without using its full capacity for every operation.

It is also natively multimodal: it can analyze text, images, documents, charts and photographs. In practice, this lets you review a technical blueprint, locate a specific part in a satellite image or extract evidence from a PDF without splitting the process across several tools.

The company says the model is especially strong at:

• Coding and analyzing complete repositories.
• Agents that can search for information, use applications and deliver results.
• Cybersecurity, from detecting vulnerabilities to creating defense rules.
• Finance, legal matters, science and engineering.
• Visual understanding of documents, scenes and technical drawings.

According to the evaluations cited by Mistral, Large 4 scores 82% on a test that requires reproducing a real software vulnerability and then fixing it. It also solves 93% of the exercises in Cybench, a set of 40 cybersecurity challenges.

These results need context. Mistral notes that some closed models reject this type of task because of their safety policies, even when they can be used to check and repair a flaw. Large 4 aims to offer that capability under the rules of the organization deploying it, including the option to run it in a private cloud or on its own servers.

Coding and tool use

In coding, the model scores 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, according to figures compiled by Artificial Analysis. Its combined score on the Coding Agent Index is 49.8%.

In a blind human evaluation conducted by Surge AI, programming professionals gave it an average score of 3.74 out of 5. It ranked second among five models, behind Claude Opus 5 and ahead of Kimi K3, GLM-5.3 and GLM-5.2.

For office tasks, it scored 59.9% on AutomationBench, which tests workflows using applications such as Gmail, Google Sheets, Slack and Salesforce. It also reached 1,393 Elo points on AA-Briefcase, a test of long-running knowledge tasks that includes documents, presentations and spreadsheets.

That points to a specific use case: an agent that gathers information from several sources, calculates results, updates a financial spreadsheet and prepares a final report, rather than simply answering questions in a chat.

A model trained and deployed in Europe

Mistral says it trained Large 4 from scratch in its own European data centers, using 3,800 NVIDIA Grace Blackwell GPUs. A significant portion of the training data was multilingual and covered more than 160 languages, including all the official languages of the European Union.

The company is also preparing an end-to-end European-operated deployment under European law. For companies and public authorities, the goal is to reduce dependence on an external provider and maintain greater control over data, usage policies and system availability.

The model was trained using the same customization and reinforcement learning environment that Mistral offers customers through Mistral Forge. Reinforcement learning allows the system to improve based on the results of its own attempts, using automated tests, judges based on other models and code checks.

What changes for you

For now, the most direct access is the preview API, with an announced price of $1.36 per million input tokens and $4.18 per million output tokens. Tokens are the small units into which the model breaks text for processing.

The important change will come when Mistral publishes the weights. If it sticks to its schedule, companies, researchers and developers will be able to run the model on their own infrastructure, adapt it to their needs and audit its behavior more freely than with a completely closed service.

That does not mean it will be easy to install on a personal computer. A one-trillion-parameter model requires substantial infrastructure, although its 49 billion active parameters may help reduce the cost of each operation compared with a model that used its full capacity at every step.

Mistral still needs to publish more details about the architecture, test results and training method. The company also warns that reinforcement learning is ongoing and expects improvements over the coming weeks and months. What matters now is whether these figures hold up in independent evaluations and whether publishing the weights really provides the control Mistral promises, especially in cybersecurity and enterprise environments.