Mistral is taking on U.S. and Chinese AI giants with its biggest model yet

Published Oct 29, 2026, 3:36 PM EDT Jefferson Mangubat is an experienced technology journalist who has reported on consumer electronics, software updates and big stories in the industry. He's now a Reporter at How-To Geek, and has built a solid reputation for delivering accurate, accessible, and insightful tech journalism. Prior to that, he was a tech reporter for Neowin, where he extensively covered hardware developments from Apple and Google, in addition to digital trends and application updates. For the past few years, Mistral has tried to carve itself a name as Europe’s answer to Silicon Valley’s AI giants, moving from smaller, open models to heavyweight systems like Mistral Large 3 in late 2025. After raising €3 billion in September 2026, the French start-up has doubled down on its ambition to compete with leading US and Chinese AI developers. Now, the company is trying again to take on the AI behemoths with Mistral Large 4, a massive open-weight model built for coding, complex reasoning, image analysis, and business automation. The model, named “Le Chonk,” is designed to provide users with greater control over their AI infrastructure, especially for more sensitive work such as cybersecurity, Mistral said in its announcement. It’s already in public preview, and the company has made its pitch for why organizations might want an alternative to closed models from US tech giants. The numbers are fittingly immense. Mistral Large 4 has roughly 1.05 trillion parameters, with 52 billion active at any given time, employing a mixture-of-experts architecture. Instead of running the whole model for every request, it routes work through a subset of the components, helping to meet the computational demands of a model this large. It also supports multi-modal input, a context window of up to 1 million tokens, and more than 160 languages. Mistral Large 4's big selling points are coding and cybersecurity Built for complex software engineering workflows Credit: Mistral Beyond just answering questions or spitting out code snippets, the company is positioning it as an AI that can handle long-running software engineering tasks, navigate repositories, use tools, and accomplish complex workflows with less hand-holding. It achieved a score of 49.8 on the Artificial Analysis Coding Agent Index, outperforming DeepSeek V4 Pro and Qwen3.8 Max in the comparison provided. In a blind coding evaluation, Surge AI finished second out of five models tested, scoring 3.74 out of five. Claude Opus 5 was top of that evaluation with 4.22. Even more interesting is the cybersecurity angle to the pitch. Mistral previously said during its preview that Large 4 got an 82% on a test where models had to reproduce and patch real software vulnerabilities, the highest score reported on that evaluation. It also solved 93% of challenges in Cybench, a set of security exercises. That ability comes with a caveat worth noting: some competing models will not do certain vulnerability-related tasks due to their safety restrictions. So their lower scores can be a reflection of policy decisions as well as technical limitations. Mistral argues that organizations need reliable access to security tools during legitimate investigations while also preventing malicious use. In Lakera’s B3 security benchmark, Large 4 defended against 93.3% of attacks. Mistral trained Large 4 on its own 3,800 European data center NVIDIA Grace Blackwell GPUs. When the model weights are available, organizations can deploy and adapt the model in their environments, subject to the model’s release license. That’s particularly attractive for industries handling sensitive information, such as finance, healthcare, manufacturing, and public services. Mistral also plans European deployments, to be operated independently under European law. Large 4 already is strong in fields like finance and legal work, scientific coding, and visual understanding, the company says. It can process complex charts, technical drawings, and lengthy documents, so it is useful for work that goes beyond the typical chatbot dialogs. Mistral Large 4 still has room to improve Le Chonk is still hitting the books Credit: Mistral Mistral is also continuing to train the model using reinforcement learning, which improves performance by using feedback on its attempts at various tasks. The company says its training process continues to yield gains, so the preview may not be representative of the final capabilities of the model. There's also the practical cost to consider. Mistral's API pricing is published at $1.36 per million input tokens and $4.18 per million output tokens, with discounted pricing during the preview period. To run a model of this size on your computer, you would also need a lot of computing power. For now, Mistral Large 4 is a promising step forward to more capable AI that businesses can control themselves. If the final model delivers on its coding, security, and reasoning claims, its open-weight release could make that proposition a lot more appealing. Source: Mistral