French AI company Mistral on Tuesday, Oct. 6, opened a preview of Mistral Large 4, its largest and most capable AI model so far. Mistral says it will make it an open-weight model by publishing the model's trained files by the end of October, so anyone can download it and run it on their own computers. Until then, developers can try the preview through Mistral Studio, the company's online service, which runs it in Mistral's own data centers in Europe.
Large 4 has 1 trillion parameters, the numbers a model tunes as it learns. It is a mixture-of-experts model, a design that switches on only part of the model for each piece of text, so it runs faster than its full size suggests: 49 billion parameters work at a time. It reads images as well as text. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell chips in its own European data centers, using text in more than 160 languages, including every official language of the European Union.
Mistral says Large 4 clearly beats every downloadable model built in the U.S. or Europe and is roughly even with the strongest downloadable models worldwide. It pitches the model hardest at cybersecurity. On one test that asks a model to recreate a real security flaw in open-source software and then fix it, Mistral says Large 4 scored 82%, the highest of any model, while leading closed models, including Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, score near zero because they refuse the task. That comparison measures willingness to do the work, not skill at it. Mistral argues that defenders need tools that will do this work, and that a company running the model on its own machines cannot lose access to it in the middle of an attack.
Until the files come out, Mistral says it is testing the model for weaknesses in real-world use with cybersecurity specialists, vetted partners and government agencies, who get the same model with fewer safety filters and wider hacking-related abilities. It says Large 4 turns down harmful hacking requests more often than any other downloadable model it compared, on three public sets of such requests.
Through Mistral's service, the preview costs $1.36 for every million tokens it reads and $4.18 for every million it writes. Tokens are the small chunks of text, often parts of words, that AI models read and write. Artificial Analysis, an independent company that tests AI models, lists the same prices. On its combined score across ten tests, the preview gets 38, against a typical 26 for comparable models, which puts it 64th of 225 in that group. It also found the model fast but very wordy: it wrote about two and a half times as much as a typical model in that group to get through the tests.
Most of the test results come from Mistral's own announcement. Mistral says it is still training the model and expects it to keep improving, so the version it releases may differ from the preview, and that it will share how the model was built, more test results and its training methods along with the files. It has not said what license the files will carry, which decides what businesses may do with them, or exactly what hardware it will take to run.