The French company says its trillion-parameter model rivals the best from the US and China, and anyone will be able to download and run it.
French AI company Mistral has released a preview of Mistral Large 4, a very large AI model nicknamed "Le Chonk". Its big selling point is that it's open-weight: businesses can download it, run it on their own computers and change it as they like.
Mistral built it mainly for cyber defence, coding and AI agents that handle long tasks on their own. The company says it's the strongest open model made outside China and close to some paid American models.
Why it matters: after the US limited access to some American models this year, many companies want AI that nobody can switch off. The full release is due on 27 October.
Mistral, the Paris-based AI company, has launched a preview of its biggest model yet. Officially called Mistral Large 4, it has picked up the nickname “Le Chonk” because of its size: around one trillion parameters, the internal settings a model learns during training.
Mistral’s pitch is simple. It says this is the most capable open model made anywhere outside China, and close to some of the paid, closed models sold by American companies. The final version, along with the files needed to run it yourself, is scheduled for 27 October.
Most of the best-known chatbots, from OpenAI and Anthropic for example, are closed. You can use them, but only through the company’s own apps or paid connections, and the company decides who gets access.
An open-weight model is different. The finished model can be downloaded, run on a company’s own servers and tuned for its own needs. You still need serious computing power to run something this large, but nobody can switch it off for you later. Running it costs only as much as the hardware and electricity it uses.
Mistral has aimed Large 4 at a few specific jobs. The headline one is cyber defence: reading through system logs, checking code for weaknesses and spotting newly published security holes. The company says it can also help find weak points the way an attacker would, so they can be fixed first.
The other big focus is agents, meaning AI that plans and carries out longer tasks on its own, such as research or coding jobs that would take a person hours. “Really, the core competency we’re looking at for ML4 is agentic capabilities,” Mistral’s vice president of science Pierre Stock told CNET.
The model also handles images as well as text. Mistral says it is especially good at reading satellite and aerial photos, for example to spot wildfires, and it was trained on more than 160 languages.
The release comes at a tense moment for AI. In June, the US government placed temporary limits on how models from OpenAI and Anthropic could be distributed, over fears they could be misused in cyberattacks, Wired reports. That has left some businesses and governments outside the US worried that access to the tools they depend on could be cut off.
Mistral is leaning into that worry. “If you use a closed model, there is no guarantee it will still be there tomorrow,” co-founder and chief scientist Guillaume Lample told Wired. Customers will be able to run the model themselves or use Mistral’s European cloud, where data stays under EU rules.
The company also says it trained Large 4 from scratch, rather than by copying the answers of bigger rival models, a practice some Chinese labs have been accused of.
Mistral is a much smaller player than the American giants. It trained this model on about 4,000 Nvidia chips over two months, according to ZDNET, while OpenAI reportedly used around 100,000 for its latest model. Early outside tests suggest Large 4 holds its own against expensive closed models on some image tasks and matches strong Chinese open models such as DeepSeek on financial work.
Final benchmark results haven’t been published yet, and Mistral expects the model to keep improving over the coming months. It also helps that the company has money behind it: in September it raised $3.3 billion, the largest funding round ever for a European tech company.
Not realistically. A model this size needs powerful data centre hardware. Most people will try it through Mistral’s own services or apps built on top of it.
Mistral says the final model and its downloadable weights arrive on 27 October. Until then it is available as a public preview.
You're reading the quick version.
OpenAI released hundreds of results toward 372 major math problems on GitHub, sparking questions about verification and what it means for the math community.
An experimental platform lets you design, test, and share games using simple text prompts, with a Unity Spark integration on the horizon.
Tab is emerging from stealth with a $300 million valuation, promising to handle everyday tasks via text on familiar apps while keeping your data private.
