Mistral opened a public preview of Mistral Large 4, its new flagship model, on October 6. Initial access is available through the API in Mistral Studio. The company plans to publish the model weights by the end of October, so anyone waiting to download the model and run it in their own environment will need to wait for the next stage.
According to Mistral, the model was built for work that combines coding, tool use and visual understanding. The company demonstrates analysis of engineering drawings, finding information in PDF documents and working with large images. These are examples presented by the developer; we have not yet independently tested the quality of the results.
The official documentation lists a context window of one million tokens, alongside support for structured outputs, function calling and document questions. For developers, this combination may be relevant to tasks that involve gathering information, invoking a tool and returning an answer in a software-readable format. Suitability for a particular workflow still requires testing with its own materials and processes.
When checked on October 7, the official pricing page showed sale prices of $0.68 per million input tokens and $2.09 per million output tokens. Cached input was priced at $0.07 per million tokens. These are API usage prices for the Standard tier shown on the page, rather than a chat subscription price. The page checked did not state when the sale ends.
Mistral also emphasizes its European infrastructure: according to the company, the model was trained in its European data centers, and the preview is served from the same infrastructure. The company presents evaluation results in coding, cybersecurity and professional work, but its published figures are no substitute for testing speed, accuracy and cost on a real task.
The immediate development is another option to try through an API. The next checkpoint will be the release of the weights and the additional architecture and training details Mistral has promised. Until then, it is worth treating this version as a preview that remains in development.
What else Mistral disclosed
According to the announcement, Mistral Large 4 is a natively multimodal Mixture-of-Experts model with 1 trillion parameters, 52 billion of them active. The documentation lists 1.05 trillion total parameters and a 1.6 billion-parameter vision encoder. It is the company’s largest and most capable model to date, and Mistral describes it as a single model that unifies instruction following, reasoning and agentic work.
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs. A significant share of the training data was multilingual, spanning more than 160 languages, including every official language of the European Union. Until the weights are released, Mistral is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.
Among the figures Mistral reports: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4, 93% of the challenges in Cybench, and 59.9% on AutomationBench, which covers 657 business workflows. On Lakera’s B3 AI Security Benchmark, the model resisted 93.3% of attacks. These are the company’s own results.
Mistral says ML4 is the first milestone on a roadmap funded by its €3 billion Series D, and that it will serve as the base for a new generation of specialized Mistral models.
What it means for you
- Developers in the US, EU and UK: the preview API is open in Mistral Studio now. Mistral says the model will be available in multiple regions worldwide but has not published a country list for the preview.
- Teams in the EU: Mistral says it will run a European deployment end to end, independently of other digital service providers and under European law. If GDPR or data-residency rules shape your vendor choices, that is the detail to check in Mistral’s terms.
- Price: the sale prices are half of the original prices shown in the documentation: $1.36 per million input tokens, $4.18 per million output tokens and $0.14 for cached input. Budget for the full price as well.
- Compare options: see more tools we have covered in our AI tools directory, and another open-weight option in our DeepSeek review.
Sources
- Mistral AI, official blog: Introducing Mistral Large 4, October 6, 2026
- Mistral AI, official documentation: Mistral Large 4 model page (pricing, context window and features), checked October 8, 2026
