Mistral AI has released Mistral Large 4 (ML4), a new large multimodal model containing one trillion parameters. French president Emmanuel Macron called this launch “a third way in AI,” positioning the system against both American and Chinese rivals.
The model is nicknamed Le Chonk. It is not an open-weight release at this stage. Users can only access it through a public guardrail endpoint. Mistral plans to publish the weights in three weeks once safety testing finishes.
“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,” said Pierre Stock, vice president of science at the company. He told TechCrunch that security worries have grown among enterprises and institutions. Stock noted that an open-weight model allows for easier auditing.
ML4 was trained entirely on Mistral’s own compute infrastructure. The lab used 4,000 Nvidia GPUs. This is two to three times fewer than Chinese competitors and significantly fewer than closed-source rivals, Stock said.
Benchmark results are not yet available. Mistral hopes the model will lead among open-weight options, particularly outside China. Focused training could also allow it to beat closed models in specific areas important to customers, where multimodal capabilities add value.
Stock identified cybersecurity and finance as key use cases. Chip design is another area, which is central to two of Mistral’s main backers. Dutch giant ASML led the company’s Series C funding round. Samsung led the Series D last month, valuing the firm at €21 billion, or about $24.39 billion.
The company previously stated that hosting Chinese models was not a pivot to becoming a mere inference provider. With Le Chonk in its corner, Mistral believes it should still be considered a frontier lab.
What it means
Developers and businesses gain a European alternative to the dominant Chinese and American stacks. The ability to audit the weights later offers a path for those needing to verify safety without relying on black-box systems. The efficiency in compute usage suggests lower costs for running the model compared to peers.
