Mistral Large 4 ‘Le Chonk’ Hits 1 Trillion Parameters, Weights Land in Weeks

Mistral has unveiled Mistral Large 4, a one-trillion-parameter multimodal model nicknamed “Le Chonk” — the French lab’s most ambitious bid yet to plant a European flag at the AI frontier, and its loudest argument for open weights. It is live in public preview now, but the full checkpoint stays behind a guarded endpoint, with the weights due in about three weeks and a release pencilled in for October 27.

The timing is pointed. Large 4 lands weeks after European Commission President Ursula von der Leyen floated a voluntary brake on the most capable AI models in her State of the Union address and invited frontier labs into the discussion. Shipping open weights now reads as a deliberate counter-position: Mistral argues that openness, not closure, is how advanced-model risk should be managed.

• Scale: One trillion total parameters, with 49 billion active per inference step via a sparse mixture-of-experts design.
• Training: Built from scratch in about two months on roughly 4,000 NVIDIA Grace Blackwell GPUs inside Mistral’s own European data centres.
• Reach: Training data spanned more than 160 languages, including every official language of the European Union.
• Licensing and funding: The checkpoint ships under a custom licence rather than Large 3’s Apache 2.0 terms, on the back of a Samsung-led Series D that valued Mistral at €21 billion (about $24.4 billion).

A trillion parameters, a fraction of them running

The headline figure describes capacity, not the cost of any single token. Large 4 activates just 49 billion of its parameters at inference, up from the 675 billion total and 41 billion active parameters of Large 3. That sparse architecture keeps serving compute in check, though the full checkpoint still requires a substantial multi-GPU deployment.

Pierre Stock, Mistral’s VP of Science, framed the training cluster as part of the pitch: roughly two to three times smaller than those used by Chinese competitors, and significantly smaller than what closed-source rivals spend. The company’s launch post cites 3,800 Blackwell GPUs, while the broader rollout is described as about 4,000.

Cybersecurity: where open weights stop being abstract

Cyber defence, coding and agentic workflows anchor the release. On the independent Artificial Analysis Cyber Index, Large 4 ranks among the top five models globally and leads open-weight models built outside China by a wide margin. On a test that asks a model to reproduce a real vulnerability in open-source software and then patch it, it scores 82% — the highest of any model tested — and clears 93% of Cybench’s 40 security-competition exercises.

Several leading closed systems, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that vulnerability task because their safety filters refuse it. Mistral’s contention is that defenders need systems capable of the work attackers already extract from jailbroken models.

“We’ll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks.” — Pierre Stock, VP of Science, Mistral

Samsung, ASML and the sovereignty pitch

Samsung led the Series D, while Dutch semiconductor equipment giant ASML led the earlier Series C. The investor mix is no accident — chip design is among Large 4’s optimised workloads, alongside finance, manufacturing and technical imagery. For European banks, hospitals and public administrations, running a frontier-class model on-premises with no data leaving their own infrastructure is a selling point closed systems cannot match on the same terms. The backdrop is a widening funding gap: European Central Bank data puts cumulative AI investment at roughly 75% in the United States against 5% in Europe.

Mistral co-founder and chief scientist Guillaume Lample argues the dividing line is control, not geography:

“Sometimes, people like to [make a big deal] over the US, versus Europe, versus China. But what really matters is to own the model — even for US companies. If you use a closed model, there is no guarantee it will still be there tomorrow.”

Competitive benchmarks, an unclosed distribution gap

Mistral reports 62% on DeepSWE v1.1, just above the 61% it lists for GLM-5.3. The live leaderboard complicates that: GLM-5.3 and Kimi K3 sit near 69% in their best published configurations, while GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 hover around 74%. Results are stronger elsewhere — a 15% task-pass rate on Harvey’s Legal Agent Benchmark and 67% on Finch, level with DeepSeek V4 Pro 0813. Mistral concedes Le Chonk still trails the frontier in some coding work.

The decisive test comes on October 27, when developers can run Le Chonk outside Mistral’s API and see how much benchmark performance survives their own workloads. The weight release, paired with Samsung’s backing, points to a hardware distribution layer Mistral has so far lacked — OpenAI commands hundreds of millions of ChatGPT users, while Mistral lives largely in developer and B2B API territory. Europe keeps building foundational infrastructure, from semiconductor investment to the IMEC centre planned in Málaga for 2027, while American firms hold the application layer. Le Chonk meaningfully narrows the technical gap and gives European enterprises an auditability and sovereignty argument closed labs cannot replicate on the same terms. Distribution is the harder problem, and the next three weeks will show whether open weights are the lever that moves it.