What happened
Mistral released the weights of Mistral Large 4, internally and affectionately called Le Chonk. It is a flagship, natively multimodal model with 1 trillion parameters, and it is fully open-weight.
This is a second act. Earlier this month Mistral put the model out in public preview. The company says the response exceeded its expectations, then folded in the feedback and kept scaling reinforcement learning. The result, per Mistral, is big gains in coding and defensive cybersecurity, with "no signs of saturation".
The pitch is aimed squarely at developers. Mistral says Le Chonk leads open-weight benchmarks in coding and defensive security. That is a vendor claim, so treat the leaderboard as something to verify, not something to frame.
There are also three reasoning modes, which let you trade cost against performance depending on the job. Quick question, cheap mode. Gnarly refactor, spend more.
Why it matters
Open-weight models at this scale are rare. A trillion parameters used to be the kind of number that stayed behind an API and a pricing page. Now the weights are out, which means teams can customize the model and run it themselves.
The focus is also telling. Agentic coding and cybersecurity are where developers are actually spending money on AI right now. Mistral is positioning an open model as a serious option for defenders, not just a hobbyist toy.
Mistral's benchmark claims cover open-weight models, not everything on the market. That is a narrower crown than "best model", and the announcement is worded accordingly.
What it means for Europe
Mistral frames this as sovereign AI: European-trained, fully open-weight, deployable on your own infrastructure. For governments, banks and any company that would rather not send sensitive code to someone else's cloud, that is the whole point.
Self-hosting means you control where the model runs and what it sees. It also means you need hardware that can hold a trillion parameters, which is a very European kind of paperwork problem and a very physical one.
A European lab shipping a frontier-scale open model is the strongest argument yet that the region is not just a customer in this race.
What to do with it
- Check the benchmarks yourself.Run Le Chonk on your own codebase and security tasks, because "leads open-weight benchmarks" is Mistral's claim, not your results.
- Test the three reasoning modes.Map each mode to a real workload and see where the cost-performance trade-off actually lands for you.
- Scope the hosting bill.Before promising "sovereign AI" to your boss, price out the infrastructure a 1-trillion-parameter model needs.
- Read the license and the weights page.Confirm what open-weight lets you do in production before you build on it.
Ciao, sapiens. The chonk is loose, and this time you can download it.


