Mistral just gave away a trillion-parameter model it says can go toe-to-toe with the best closed systems from the US and China.
Mistral Large 4, which the company is calling Le Chonk, is a 1 trillion-parameter model now available in preview, with a final release due by the end of October. It's free for anyone to use and customize. Rather than chasing general benchmarks the way most frontier labs do, Mistral is leaning into coding and cyberdefense, plus specialized tasks in manufacturing, finance, and electrical engineering. Chief scientist Guillaume Lample frames the bet as going after areas bigger labs skip: "There are a lot of areas where the other labs will not focus that much."
That's a pointed contrast to the access fights playing out between US and Chinese AI labs, where the strongest models stay locked behind APIs, licensing deals, and export restrictions. By handing out a trillion-parameter model that anyone can fine-tune for a specific industry, Mistral is betting that open weights plus narrow specialization can beat closed, general-purpose scale for customers who just want a model that's good at their job.
It's a smaller, scrappier strategy than trying to out-benchmark GPT or Gemini - and if Lample is right, the niches may end up mattering more than the leaderboard.