Mistral previews 1T-parameter Large 4 ahead of weight release
The Paris-based company has opened API access to its new model and says its weights will be available on October 27. Benchmark results remain preliminary, with reinforcement learning still under way.
Mistral says its 1T-parameter Large 4 is the most powerful open-weight AI system outside China. The Paris-based company has released a preview through its API, while model weights are due on October 27. That distinction matters because API access permits use through Mistral’s service, whereas weights let users download and run the model themselves.
The preview marks a release in stages, rather than immediate public availability of the weights. Mistral says developers can access the model through the API now, and that weights enabling local operation will follow. No regulator, investigation, legal decision or adopted rule is identified in the supplied material. The company’s stated position is that the system is powerful by the comparison it makes, although its reported benchmarks are preliminary.
API access precedes downloadable weights
For developers building on regulated technology, the immediate access route is the API. This allows them to send work to the model through Mistral’s service, without receiving the weights for running it on their own servers. The source does not state contractual terms, access controls, data handling conditions or geographic restrictions. Those points therefore cannot be assessed from the release information provided.
The weights have a different practical significance because they allow users to download and operate the model independently. Mistral says they will be available on October 27, but the source does not specify a licence or any conditions attached to that release. The company has not stated in the supplied text how it will respond to particular regulatory or legal requirements. The release date is presented as a company commitment, not as a date set by an authority.
Training details and early benchmark results
Mistral says Large 4, also called “Le Chonk”, can process text and images. It uses 49 billion parameters for any task, although the model has 1T parameters in total. The company says it trained the model from scratch over two months using about 4,000 Nvidia Grace Blackwell GPUs in its European data centres, and that training used about 10 megawatts of power.
The reported performance remains provisional while reinforcement learning continues. Mistral describes that phase as improvement through trial and error and feedback, and says benchmark results may change before weights are released. On DeepSWE, an agentic coding test involving software engineering tasks, the company reports a score of 62%. It also reports 67% on FinWorkBench and 15% on Harvey Legal Agent Benchmark; the supplied source says low scores across that legal test show models remain far from handling legal work independently.
Company claims and unresolved release terms
Mistral reports a 73% score on DIOR-RSVG for visual grounding, which means locating particular objects within images. It reports 68% for GPT-6 Astra on the same test, and says both systems scored 42% on Dense200. The company also claims that Large 4 outperforms Anthropic Fable on these tasks, but has not shared figures for that comparison. These are company-reported preliminary results, not an assessment by a regulator or an independently established legal finding.
The source gives no regulator’s response, legal status, licence text or specific compliance commitments. It also leaves open whether the reported results will change before the weights are made available. Mistral says it will test the model under real-world conditions over the next three weeks, with developers, cybersecurity firms and government agencies receiving a less restricted version with additional cyber features. The next concrete step stated by the company is the planned weight release on October 27.