Mistral pitches its trillion-parameter Le Chonk to teams that want to own their model

The customer Guillaume Lample has in mind is a security team whose defenses run on somebody else's model. According to Lample, Mistral's cofounder and chief scientist, a business that relies on a proprietary model to help fight off cyberthreats risks the sudden collapse of its defenses ^[11]. "If you use a closed model, there is no guarantee it will still be there tomorrow," he said ^[10].

Buyers of US models have reason to take that seriously. On top of the June distribution limits ^[14], the White House is reported to have asked US labs to keep unreleased models away from even the UK's AI Safety Institute ^[15]. Lample aims the argument at American buyers as well as European ones. He said that "what really matters is to own the model," even for US companies ^[9].

The pitch is ownership. Anyone can use and customise Large 4 ^[1]. For a team without its own cluster, the realistic purchase is metered access from Mistral. The company earns pay-as-you-go fees for running models on its cloud and sends engineers to tune models for customers ^[13]. Its earnings have reportedly risen 20-fold in the last year or so ^[19]. A team renting Le Chonk that way still has a vendor. The difference is that the weights are open, so the team can take the model elsewhere if the terms change. Open-weight models cost only as much as the compute they consume, WIRED notes ^[12], and a team hosting a trillion-parameter model itself pays for all of that compute ^[1]. WIRED's report does not include hardware requirements or benchmark scores.

On quality, the evidence so far comes from Mistral. The company presents Le Chonk as by far the most capable open-weight model built outside China and "very, very close" to some proprietary models ^[17]. It says the release will remove the few remaining reasons a business might hesitate to choose open source ^[18]. WIRED's description of the company's record is less flattering: with less capital and compute than OpenAI and Anthropic, Mistral has generally lagged on model performance, revenue and release frequency ^[8]. Speaking to WIRED in July, Andrea Renda, research director at the Centre for European Policy Studies, said EU sovereignty goals and US hostility put Mistral, "whose performance has not been spectacular," in "a favorable position" ^[16].

Lample's case for the niche tuning is that the bigger labs leave room. "There are a lot of areas where the other labs will not focus that much," he said ^[6]. A team doing general office drafting gets less from that tuning than one doing electrical engineering or finance work ^[5].

Two questions sort the decision for whoever has to make it this week. The first is what breaks if your current model is withdrawn next week. The second is whether your work sits in the domains Mistral tuned for. A security team running triage on a closed US model scores high on both. It has the strongest case to test the preview against its own past incidents now, and to move production only after the final version lands by the end of the month ^[2]. Where exposure is high but the work is general, the open weights are insurance, and the test is whether quality on your own tasks is close enough to live with. Niche work with little exposure can be judged on output and compute cost alone. Teams low on both will find nothing in this release that requires a move this month.