Mistral Large 4 ('Le Chonk'): the French giant aiming to redirect open AI

French startup Mistral AI launched Mistral Large 4 (ML4) — unofficially nicknamed "Le Chonk" — on Tuesday, October 6, a 1-trillion-total-parameter, natively multimodal language model with 49 billion active parameters. It is the largest and most capable model the company has produced to date, and its announcement carries implications that extend far beyond benchmark scores.

Why "Le Chonk"?

The nickname "Le Chonk" originated as an inside joke within the open AI community. "Chonk" — internet slang for an exceptionally wide cat — first appeared as the "Le Chaton Fat" meme when fans were pressuring the company to build a truly massive frontier model. The team embraced the joke: "Le Chaton Fat" never existed as a real model, but now Mistral itself is launching a real 1-trillion-parameter model with a fat-cat codename. "The joke therefore lands unusually close to reality," as VentureBeat put it after interviewing Guillaume Lample, co-founder and chief scientist.

This convergence between meme and product is more than marketing flair. It signals something deeper about how the open-source community has been shaping — or at least pressuring — the trajectory of AI innovation. Public demand for increasingly large models has created a feedback loop that Mistral itself acknowledges has influenced its own roadmap.

Architecture: 1 trillion parameters, but only 49 billion active

ML4 is a natively multimodal model (text and image input, text output) built on a Mixture-of-Experts (MoE) architecture. Although the model has 1 trillion total parameters, only 49 billion are activated for each inference operation — drastically reducing compute cost and latency compared to dense models of the same scale. Think of it as a football squad of 1,000 players with only 49 on the pitch at any given moment. The result is a model that scales in capacity without scaling linearly in cost.

Training was done from scratch over roughly two months on 4,000 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. The model was trained across more than 160 languages, including all official EU languages. This infrastructure choice is deliberate: Mistral is positioning ML4 as a solution that can run under European sovereignty, without depending on non-EU cloud providers.

Benchmarks: strong numbers, with caveats

Preliminary results are impressive on several metrics. ML4 scored 62% on the DeepSWE v1.1 benchmark (a long-horizon software engineering evaluation), outperforming Reflection's Beam (44%), Qwen 3.8 Max (51%), and DeepSeek V4 Pro (57%). On the Artificial Analysis Cyber Index — an independent evaluation of how well models find and fix security flaws — the model ranks among the top five globally and leads open-weight models developed outside China.

In a specific security test where the model must reproduce a real vulnerability in open-source software and then patch it, ML4 scored 82%, the highest reported score. It also solved 93% of challenges in Cybench (40 exercises drawn from security competitions).

However, there are nuances. VentureBeat noted that Mistral's results do not yet appear on public leaderboards (Artificial Analysis, DeepSWE), and some benchmarks cited in Mistral's slides were not found in independent public sources. Mistral's ranking claims remain provisional until external researchers can test the final model and the weights are publicly released.

The geopolitical context: Europe's third way

ML4's launch does not occur in a vacuum. China has Kimi K3 (from Moonshot AI), a 2.8-trillion-parameter model launched in July. The US has Claude, GPT-6 Astra, and other closed models from OpenAI, Google, and Anthropic. Europe, meanwhile, is attempting to build a "third way" — open models, sovereign, running on European infrastructure.

This is not just continental pride. It has practical implications. For sensitive sectors like cybersecurity, defense, and government, data sovereignty is not a luxury — it's a requirement. If a closed model can be remotely shut down (as has already happened with some providers), the interruption can be a direct operational risk in critical incidents.

Mistral knows this. That's why the model is being released in preview to developers and "cybersecurity leaders" before the general weight release. For three weeks, government authorities and vetted partners will access the same model, "with reduced moderation and expanded cyber capabilities." The company says it will not interfere if its AI is used for defense — a clear signal that it is targeting government contracts.

The business bet: open weights + sovereign stack

What makes ML4 different from other frontier models is its distribution strategy. Models like Claude or GPT are "closed" — only accessible via API, with moderation controls and no possibility of local customization. Chinese models like Qwen and Kimi are open, but European sovereignty demands European infrastructure.

Mistral is betting on the middle ground: open weights (to be released October 27) + European infrastructure (own data centers) + complete enterprise stack. This means clients can not only run the model locally but also customize and integrate it with existing systems — essential for banks, factories, and governments.

In September, the company announced a €3 billion Series D round at a post-money valuation above €21 billion — the largest equity fundraising ever completed by a European technology company. Part of that capital is already being deployed to scale compute capacity in European data centers. The science team has grown from three researchers to roughly 300.

What to expect next

ML4 is described by Mistral itself as "the first milestone" on a roadmap funded by the Series D. That means more models will follow — specialized, optimized, built on ML4's foundation. The company has already confirmed it is continuously refining the model during the preview period using reinforcement learning to improve the final checkpoint.

Whether Europe's bet on open, sovereign AI can achieve commercial scale sufficient to rival the massive investments of OpenAI, Google, and Anthropic remains an open question. The next six months will be decisive: when the weights are released, when the first government contract is signed, and when the first autonomous agents based on ML4 begin running in production.

Sources: Wired, VentureBeat, The Hindu

✓ Independent sources cross-checked and verified before publishing