Mistral Large 4 is on the API now, open weights due end of October

Mistral AI, the French lab, has released Mistral Large 4 through its API. The company says the open weights will follow at the end of October. If you run models at home, that second date matters more. You can call the model today, but you can’t run it locally yet.

Mistral’s launch post calls the model “Le Chonk.” It lists 1 trillion parameters, native multimodality, and 49B active parameters. The sources don’t explain how the active count works inside the model, so treat that figure as Mistral’s own description. The company also says it is the best open-weights model from the US or Europe on aggregated benchmarks. The claim rests on averaged benchmark scores. No independent ranking is cited.

The tests Mistral chose lean toward security and agents. In a CTF speedrun, a timed set of hacking puzzles, the model solved 18 of 19 challenges. In a malware test, it got an unknown binary and identified it as Cobalt Strike, a known attack framework. It then pulled out the indicators of compromise (IoCs) and the malware configuration. On AutomationBench, a set of 657 business workflows across simulated apps like Gmail, Google Sheets, Slack and Salesforce, Mistral says the model finished ahead of several rivals. On a legal agent benchmark (LAB), it ranked first among open-source models.

Cybersecurity partners are testing the model privately, so the security numbers come from Mistral’s own runs for now. The pitch is agents that gather information and produce finished deliverables across complex workflows. Whether the model fits on a home GPU isn’t covered. One reply joked about 12 GB of VRAM. The sources say nothing more on that.