Mistral's trillion-parameter 'Le Chonk' is real. Its benchmark lead is not.

Mistral launched its trillion-parameter flagship on Tuesday, and the joke wrote itself. In June, a fictional model called "Le Chaton Fat" went viral, with fake benchmark charts for a nonexistent giant French model. CEO Arthur Mensch played along. Now Mistral Large 4 is real, codenamed "Le Chonk," and it is a one-trillion-parameter sparse mixture-of-experts model with 49 billion active parameters, trained over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in the company's own European data centers.

The model is in API preview now. The weights arrive October 27 under a custom Mistral license, after a three-week preview period with developers, cybersecurity leaders, and government authorities. Mistral says it will keep doing reinforcement learning and tune the final checkpoint in the meantime. Take that as a warning label on every number below: these are company-supplied results for a checkpoint that is still moving.

Mistral's claim is that Large 4 is the strongest open-weight model developed outside China and competitive with the best Chinese open systems. The charts it supplied show 62% on DeepSWE v1.1, a long-horizon software engineering benchmark, ahead of DeepSeek V4 Pro at 57% and GLM-5.3 at 61%. On Harvey's Legal Agent Benchmark it claims a 15% task-pass rate, ahead of the open models on the public Vals.ai board. Its chart also claims 67% on Finch, an enterprise finance workflow benchmark, tied with DeepSeek V4 Pro. Note the verb: these are all Mistral's numbers.

Some of those competitor figures check out. VentureBeat traced Reflection's own Beam announcement to matching numbers for Beam and Qwen 3.8 Max, and Artificial Analysis to the 57% for DeepSeek V4 Pro. The Harvey legal numbers match the public Vals.ai leaderboard. So the charts are not fabricated. They are selective.

Here is the rest of the picture. The live DeepSWE leaderboard, selecting each model's best published configuration, puts GLM-5.3 and Kimi K3 around 69%, and closed models from OpenAI, Google, and Anthropic around 74%. Mistral's 62% is competitive against Western open-weight models and does not establish an outright coding lead. The Finch scores for the newer models in Mistral's chart could not be independently located by VentureBeat. The visual-grounding competitor scores in the slides do not appear in the public benchmark sources reviewed. And Large 4 does not appear in Artificial Analysis' public evaluations or on the DeepSWE leaderboard at all. The ranking claim is provisional until outsiders can test the final model and the weights. As of this morning, nobody outside Mistral can.

The interesting part of this launch is not the leaderboard. It is the pitch. Mistral is positioning Large 4 as infrastructure that enterprises and governments run themselves: sovereign deployment, zero-data-retention options, continued customization. The company argues that security teams cannot depend entirely on closed providers whose safety systems might refuse dual-use but legitimate defensive requests. That argument has a real buyer. It is also worth reading the other way: an open-weight model strong enough for cyber defense under a customer's own policies is strong enough for cyber offense once the weights are public and the safeguards are stripped. Mistral says it is red-teaming with cybersecurity leaders, vetted partners, and state authorities, who get a version with reduced moderation and expanded cyber capabilities. The preview gating is doing genuine safety work here, or at least buying time for it.

The company's context matters for reading all of this. Mistral closed a 3 billion euro Series D in September at a post-money valuation above 21 billion euros, which it described as the largest equity round ever completed by a European tech company. It says it serves more than 125 global enterprises. The strategy is explicit: model weights commoditize, and the enterprise business moves to the systems around them. That is why releasing powerful weights is comfortable for Mistral and would be terrifying for a company whose moat is the weights alone.

For operators choosing models today, the honest summary is short. Large 4 does not change the open-weight decision on October 7. The weights are not out, the rankings are provisional, and the numbers that matter will be the independent ones. But it does change the vendor conversation: Mistral is now the company betting hardest that control, not just capability, is the product. If your deployment needs sovereign infrastructure or your security workflows keep hitting a closed provider's refusal policy, the October 27 checkpoint is the one to evaluate. Judge it then, on someone else's hardware, against someone else's benchmark. The meme was funny. The due diligence still has to be boring.

We have been tracking the open-weight race alongside the pricing war on the closed side (see our rundown of September's coding-agent price cuts). The pattern is the same in both: capability gets cheaper, control gets more valuable. Mistral is selling the second one now.