French artificial intelligence champion Mistral AI on Tuesday launched a public preview of a new frontier model that it says will be the most capable open-weight model developed outside China once its weights are released.
The launch comes as the race over open-weight AI, in which a model’s parameters can be downloaded and run by anyone, has taken on a geopolitical edge.
Chinese developers have set the pace, while many American companies, such as OpenAI, keep their most capable systems closed. Meanwhile, the debate over safety has grown since AI agents have been found to go rogue and hack websites.
Mistral is betting that its new open-weight model, called ML4, which has strong defence capabilities, will attract customers who want a European alternative built around “sovereignty.”
Here is everything to know about the model.
ML4, nicknamed “Le Chonk,” meaning the large chunky one, is “a new generation of models,” Guillaume Lample, chief scientist and co-founder of Mistral, said in a press conference that Euronews attended.
The model focuses on cybersecurity and defence capabilities, which the company claims are stronger than those of its Chinese competitors.
“The cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks,” said Lample.
Mistral’s launch comes just a day after Reflection, a US startup, unveiled Beam as its first open-weight model and a Western counterweight to Chinese open models. Its weights are due later this month, and its performance claims have not been independently verified.
The rogue AI agents
Pierre Stock, Mistral’s first employee and vice president of science, said that building a new AI model was “like building a spaceship”.
He said every AI company is “moving at full speed”. Though Mistral is only three years old, it is “definitely closing the gap” with its competitors, and in some aspects its new model is “basically stronger than China’s models from this summer,” he said.
Last month, the company announced it raised €3 billion at a valuation of more than €21 billion, which it says is the largest equity round ever completed by a European technology company.
The model’s release also comes as AI companies report AI agents acting on their own and hacking websites.
Stock argued that this has created a narrative for US companies to “market rogue events” as “only a few trusted players can take care of the technology.”
He added that the “monopoly is controlled by the Americans but that open-weight models allow people to access models.”
The company also said it leads open-weight models on manufacturing and finance tasks and on multimodal use, and that in coding it is closing the gap with other frontier models. Mistral also claimed the best results of any model — closed ones included — on grounding capabilities, and top-level performance on semiconductor benchmarks.
The French company said ML4 will get better, as training is not finished and reinforcement learning will conclude in the coming days.
Made in Europe
Mistral said ML4 was trained from scratch over two months on 4,000 NVIDIA Grace Blackwell GPUs in its own European data centres.
ML4 covers more than 160 languages, including every official EU language, and works across Latin and non-Latin scripts.
When will it be available?
The model is now available through Mistral’s API on Mistral Studio. Mistral plans to release the weights on October 27.
Until then, developers, cybersecurity leaders and state authorities will be able to test the model in real-world settings.
