The strongest argument for open-weight AI has never been price. It’s that nobody can revoke your access. Mistral is leaning hard on that point with Mistral Large 4, a trillion-parameter model nicknamed “Le Chonk,” whose weights the company announced are now available on Hugging Face and through its API.
The model was introduced earlier this month. What changed is that businesses and governments can now download it, run it on their own hardware, and keep it. Weights are the numerical parameters learned in training, so having them means having the working model, not a rented endpoint.
Mistral is aiming this squarely at cybersecurity. That’s a smart choice. Security teams are the buyers most likely to balk at sending logs, source code and network data to OpenAI or Anthropic, and on-prem deployment answers that objection directly. Pierre Stock, Mistral’s VP of Science, told TechCrunch that access to the model “can’t be cut off by definition because the model is open-weighted.” For a government agency worried about vendor dependence, that line is the whole sales pitch.
On performance, Mistral says ML4 is comparable to GLM-5.3, the Chinese open-weight model that was widely seen as the best of its kind for cyber work. The difference is cost: Mistral claims ML4 uses 47 percent fewer tokens per task. If that holds up outside Mistral’s own testing, it matters, because token usage drives inference bills for agentic workflows. The company also says ML4 ranks first among open models on Artificial Analysis’ Coding Agent Index.
Treat those numbers as claims until independent evaluations land. Vendor benchmarks on cyber tasks are notoriously selective. Still, the GLM-5.3 comparison is telling. Mistral is positioning itself as the Western, and specifically European, alternative to Chinese open models, which many governments hesitate to adopt. That “sovereign” framing is the company’s real differentiator against both GLM and the US closed labs.
Stock also made a defensive case for openness. AI is helping security researchers find bugs faster, he argued, but attackers are using the same tools, so defenders should get the capability “as soon as possible.” It’s a familiar argument, and critics will point out that releasing a trillion-parameter model with strong cyber skills helps attackers too. Anyone can download weights, and there’s no kill switch. Stock acknowledged the risks around rogue AI agents but said, “We don’t want to fearmonger.”
Does it matter? For enterprises with strict data rules and the GPUs to host a model this size, yes. A trillion parameters isn’t cheap to serve, so most small teams will stick with APIs. But for governments and large security vendors, ML4 is now a credible, self-hostable option, and it puts real pressure on closed providers to justify their dependence model.
