Mistral AI has released Mistral Large 4 (ML4) in public preview. The new model has 1 trillion parameters, 49 billion of which are active. The weights are not yet available, Mistral plans to release them on October 27.
ML4 is a multimodal Mixture-of-Experts model capable of processing text and images, and is focused, among other things, on programming and performing tasks via AI agents. The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s data centers across Europe. The current API preview also runs on that infrastructure.
The training took about two months and consumed approximately 10 MW of power, reports TNW. Furthermore, the reinforcement learning phase is still ongoing. According to Mistral, the improvements have not yet plateaued. As a result, the preview’s benchmark results are preliminary and may still change before the weights are made available.
Strong claims regarding cybersecurity
Notably, Mistral places a strong emphasis on cybersecurity with ML4. The company argues that security filters in closed models sometimes also prevent legitimate research into vulnerabilities. With a self-hosted model, organizations can determine for themselves which restrictions to apply in this regard.
Mistral supports this positioning with various benchmarks. According to the company, ML4 achieves an 82 percent score on a test designed to reproduce and subsequently patch an existing vulnerability in open-source software. On Cybench, which includes forty security tasks, the score is 93 percent.
However, these results should be viewed with some caution. Scores from different models are not always directly comparable. For example, Mistral points out that some closed models score poorly on certain security tests because they refuse tasks, and not necessarily because they are technically incapable of performing them.
Other results also need to be put into perspective. On Harvey’s Legal Agent benchmark, for example, ML4 achieves 15 percent, compared to 13 percent for Kimi K3 and 5 percent for GPT-6 Astra, reports TNW. While ML4 holds a lead, the low scores indicate that independent legal work remains problematic for these models.
European infrastructure
Mistral also emphasizes ML4’s European origins. Once the weights are released, organizations will be able to run the model themselves, for example, on-premises or in a private cloud. The company links this to the discussion about dependence on U.S. AI providers.
For now, ML4 is only available via Mistral’s API. Before the weights are made public, the company is having the model further tested by cybersecurity specialists, partners, and government agencies.
ML4 is the first new model since the 3-billion-euro funding round that Mistral completed in September. That funding is being used, among other things, to expand its own European computing infrastructure. According to TNW, new capacity is expected to become available by the first half of 2027. Ahead of the release of the weights on October 27, Mistral also plans to publish more details about the architecture, benchmarks, post-training, and security tests.