The Pulse
Mistral Unveils Large 4, a 1.05 Trillion-Parameter Model
Mistral has opened a public API preview of Large 4, a 1.05-trillion-parameter model, while planning to release its weights on October 27. The company says it will publish further details on the model before the weights become available.
AI.info Team · Written with AI ·
Mistral says its new Large 4 is the strongest open-weight model from the United States or Europe, but the company is not releasing its weights yet. Announced on October 6, the 1.05-trillion-parameter model is available first as a public preview through Mistral’s API; the company plans to publish the weights on October 27, after further training and safety work. That gap puts the launch’s central tension in plain view: customers are being offered a powerful model, while broad access to run it independently waits.
1.05 trillion parameters, 49 billion active
Mistral’s model documentation describes Large 4 as a multimodal, general-purpose model built with a Mixture-of-Experts architecture. It has 1.05 trillion parameters in total, of which 49 billion are active for a given task, and a 1.6-billion-parameter vision encoder. The model page lists a context window of one million tokens.
The company calls the model “Le Chonk” and says it can process text and images. According to The Next Web’s launch-day report, Mistral trained it from scratch for two months on about 4,000 Nvidia Grace Blackwell GPUs at its European data centers. The reported training figure is about 10 megawatts of power.
Early scores come with a caveat
Large 4’s initial benchmark results put it close to leading open models in some tasks, but Mistral’s training and refinement are not finished. On the DeepSWE agentic coding test, the preview scored 62%, compared with 61% for Zhipu’s GLM-5.3 and 57% for DeepSeek-V4-Pro, according to The Next Web. On FinWorkBench, it scored 67%, level with DeepSeek-V4-Pro and just ahead of GLM-5.3 at 65%.
The same report says the model scored 15% on Harvey’s Legal Agent benchmark, where OpenAI’s GPT-6 Astra scored 5% and Kimi K3 scored 13%. Those results are not evidence that the model can handle legal work independently: scores were low across the field. Mistral also describes the benchmark results as preliminary, and expects them to change before the weights arrive.
Mistral presents Large 4 as competitive with the best open models, not as a match for the leading closed systems. Pierre Stock, Mistral’s vice president of science, told Axios that the company believes the model is the world’s best open-weight model.
“We're not there yet on the frontier,” Pierre Stock, Mistral’s vice president of science, told Axios.
Cybersecurity access is split
For now, the public API has moderation limits. Mistral is also sharing a version with fewer restrictions and broader cybersecurity capabilities with selected partners, including developers, cybersecurity firms and government agencies, for testing. The company says it plans to use the time before the October 27 weights release for additional reinforcement learning and safety testing.
The tradeoff is familiar in open-weight releases: downloadable weights give customers more control over where and how they run a model, but make it harder for the developer to control how others use it. Guillaume Lample, Mistral’s co-founder and chief scientist, told The Next Web:
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”
That defense case sits alongside the risk of releasing a capable model with safeguards that users can alter. Mistral’s staged release keeps the broad public preview behind its API while partners test the less restricted version. The next concrete milestone is October 27, when the company says it will publish the weights. It plans to publish additional information about the model’s training, benchmarks and safety testing before that release.