Mistral AI has released Mistral Large 4, a one trillion parameter model currently accessible through a public guardrail endpoint. The company says it plans to make the model’s weights available in three weeks, after safety testing. That schedule leaves users with access to the service before they can examine or run the weights themselves.
How can users access the model now?
At present, Mistral Large 4 is available through a public guardrail endpoint. A guardrail endpoint provides controlled access to a model, rather than distributing the underlying weights for users to run themselves. The source does not describe the endpoint’s specific safeguards or access conditions. Mistral AI says the weights are planned for release once safety testing is complete.
Why are the weights not available yet?
The company has linked the planned release to further safety work. Pierre Stock, Mistral AI’s vice-president of science, told TechCrunch that the company would work with trusted partners and governments in the meantime. He said that work is intended to help ensure the open-source weights can support defence without enabling malicious attacks. The source does not specify the tests, partners or governments involved.
What happened before this release?
Mistral Large 4, also referred to as ML4, is described as a multimodal model with one trillion parameters. A parameter is a value used by a machine-learning model when processing information; multimodal models work across more than one kind of input or output. The source calls the model Le Chonk, a nickname tied to its size. Mistral AI has not yet published benchmark results for the model.
What is known about its training?
Pierre Stock said Mistral Large 4 was trained entirely on Mistral AI’s own computing resources. The training used 4,000 NVIDIA GPUs, according to Stock. He said that was fewer GPUs than Chinese competitors used, but the comparison and its methodology are not detailed in the source. No training duration, energy use or cost is provided.
Which use cases does Mistral AI identify?
Stock identified cybersecurity, finance and chip design as areas suited to the model’s training. He said multimodal capabilities could be valuable for customers in those fields. These are company expectations rather than reported benchmark outcomes, which remain pending. The source gives no customer deployments or performance measurements for any of the use cases.
What remains unclear after the announcement?
Mistral AI has not supplied benchmark results, and the source does not confirm a final date for making weights available. It also does not describe the safety evaluation or how the weights would be distributed. ASML and Samsung are identified as backers, and the source links chip design to their business interests. It says Samsung led a Series D at a €21 billion valuation, while ASML led a Series C.
The next reported step is completion of safety testing before the planned weight release. Until then, access remains through the public endpoint. The release timing and evaluation details are still company plans rather than completed steps.