Mistral has released a new artificial intelligence model with an ambitious goal: make Le Chonk the strongest open-weight model available outside China. The model, officially called Mistral Large 4, or ML4, contains 1 trillion parameters and is designed to compete with both open and closed systems from the United States and China.
Mistral says ML4 is the most capable open model outside China and comes close to some proprietary models. The company also expects it to rank among the top open-weight models worldwide on combined benchmark results, calling it the strongest open-weight model developed outside China by a substantial margin.
A trillion-parameter model built in Europe
ML4 was trained for two months in Mistral’s data centers in Europe. The training used 4,000 Nvidia GPUs, and Mistral says the entire process relied on its own computing resources.
That hardware count is central to Mistral’s pitch. The company says it used two to three times fewer GPUs than Chinese competitors and far fewer than closed-source competitors. The message is clear: Mistral wants ML4 judged not only by its size, but also by what it achieved with a smaller amount of computing power.
Guillaume Lample, Mistral’s cofounder and chief scientist, said the model still has room to grow. “There are so many domains in which you can improve models,” he said. He added, “The model capabilities will further improve as we scale up our training capacity, following our Series D fundraise.”
That fundraise gave Mistral 3 billion euros, or $3.4 billion, in September and valued the company at 21 billion euros, about $24.39 billion. Mistral’s earnings have also increased 20-fold in the last year or so, giving the company more resources as it expands its training capacity.
Strong claims, with limits still visible
Mistral’s claims place ML4 near the top of the open-weight field, but the company is not presenting it as the best model in every task. ML4 still lags behind the frontier in areas such as coding, which leaves room for competing models to hold an advantage in important uses.
That distinction matters because benchmark leadership does not always translate into the same results across every task. Mistral says ML4 will rank among the leading open-weight models on aggregate performance, while also acknowledging that the model remains behind the frontier in coding.
The company is positioning ML4 as an alternative to models from both the United States and China, including open and closed systems. That gives the release a wider purpose than adding another model to the market: Mistral wants organizations to have a powerful option that does not depend on either side of that divide.
Safety testing comes before open release
Mistral plans to release ML4’s core parameters as open-source weights three weeks after safety testing is complete. The delay gives the company time to assess how the model can be used before making its weights available.
Mistral also says ML4’s cyber defense capabilities will help enterprises and governments defend against threat actors that jailbreak closed models. In this context, jailbreaking means finding ways around the safeguards built into a closed system.
Opening the model’s weights creates a responsibility that Mistral says it plans to share with trusted partners and governments. Pierre Stock, Mistral’s vice president of science, said, “In the meantime, we’ll work with trusted partners and governments to make sure that the open-source weights can be used to defend, but not to [perform] malicious attacks.”
That approach makes the three-week safety period an important part of the launch, not a footnote. Mistral wants developers and institutions to gain access to ML4, while also working to prevent the model from becoming a tool for malicious attacks.
Le Chonk therefore arrives with two promises: a 1-trillion-parameter model that can challenge the strongest open offerings outside China, and a path toward wider access built around safety checks. Its performance in coding and other frontier tasks will determine how far those promises hold as the model reaches more users.
Based on
