Mistral AI has released the weights of Mistral Large 4, the French lab's 1-trillion-parameter flagship nicknamed "Le Chonk," three weeks after the model entered public preview. The announcement, published on Mistral's blog under the title "We release the weights of Le Chonk," says the improved model now leads open-weight benchmarks in agentic coding and defensive cybersecurity, and positions the release as a milestone for European "sovereign AI."
The move converts October's most talked-about preview into a fully downloadable artifact, with the weights live on Hugging Face under the repository name Mistral-Large-4-1T-A52B. For developers tracking open AI models and the shifting balance between Chinese, American, and European labs, it is the most consequential open-weight release from a Western lab this quarter.
From preview to open weights in three weeks
Mistral launched Large 4 in public preview earlier this month, promising that full weights would follow by the end of the month — a promise the company has now kept. In the interim, Mistral says it put community feedback to work while continuing to scale reinforcement learning, driving what it describes as massive improvements in coding and defensive cybersecurity. The blog adds, notably, that "gains continue to accelerate, with no signs of saturation" — an unusual admission that the model was released with headroom still on the table.
The result is a natively multimodal, trillion-parameter sparse model that Mistral bills as the best open-weight model for coding and among the world's most capable open models overall.
The benchmark case: leading FrontierCode 2.0
The headline number comes from Cognition's private FrontierCode 2.0 benchmark, which assesses end-to-end code quality — specifically, whether a repository maintainer would merge the agent's change. Mistral Large 4 leads the field with a score of 19.7 percent, ahead of DeepSeek's V4.1 Flash at 19.3 percent, GLM-5.3 at 17.8 percent, Kimi K3 at 17.3 percent, GLM-5.3 Flash at 14.2 percent, and Mistral's own Beam 1.0 at 2.9 percent.
Those margins are thin at the top — a 0.4-point lead over DeepSeek is within the noise of any single private benchmark — but the symbolic shift is larger. Chinese open-weight models have held the top of agentic coding leaderboards for most of the fall; a European lab taking the lead spot, on a benchmark owned by an American company, reframes the competition as three-sided rather than bilateral.
Token efficiency: 47 percent fewer tokens per task
Mistral's second claim may matter more for real deployments. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48 percent — effectively tied with GLM-5.3, the leading open-weight model — while using 47 percent fewer tokens per task on average. Because agentic coding burns tokens on planning, tool calls, and retries, token efficiency translates directly into inference cost at scale, which is why the company is leaning on the word "economical."
The release also ships with three reasoning modes, letting operators trade response quality against cost per query — a recognition that most enterprise coding workloads do not need maximum reasoning depth on every call.
The sovereign AI pitch
Beyond benchmarks, the announcement is a positioning document. Mistral emphasizes that Le Chonk was "trained from scratch on European infrastructure with clear lineage from inception to release," giving organizations a trusted foundation for what the company calls Sovereign AI — models that institutions can own, customize, and self-host without depending on American or Chinese providers. The blog's tagline, "Built for Coders, Trusted by Defenders," telegraphs the intended buyers: software teams and cybersecurity organizations, the latter drawn by the model's defensive-security training.
The framing did not stop skeptics from noting the gap that remains. The Decoder characterized the release as a strong coding model that still trails China's best overall, and the benchmark table itself shows DeepSeek within half a point. Still, the community response was emphatic: the release thread on Hacker News gathered roughly 585 points within hours, and the Hugging Face repository went live the same day.
Third-party readings
The trade press largely accepted Mistral's coding claims while keeping the scoreboard in view. Fortune reported the company's characterization of Large 4 as the best open model for agentic coding and cyber defense. VentureBeat called it the top open model outside China and highlighted a 63 percent lower cost-per-task against rival open models. The Next Web noted simply that the weights of the 1-trillion-parameter model are now public, and ZDNET observed that Le Chonk now ranks alongside Codex GPT-6.1 Terra among top open coding models.
For Mistral, the open-weights bet is now explicit at trillion-parameter scale. The company releases its best models openly and sells the surrounding platform — and with Large 4, it is betting that European sovereignty, token efficiency, and a top-two open-weight coding score are enough to carve out a durable position between the American frontier labs and China's open-weight powerhouses.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
