Mistral Is SO Back: Inside Le Chonk, the West’s Biggest Open Model (Pierre Stock, VP Science, Mistral)

Full analysis

Mistral says its new open-weight model, Mistral Large 4 (nicknamed "Le Chonk"), matches the best closed models on the hardest tasks while running on roughly a quarter of the compute China's DeepSeek used, all trained on Mistral's own European data centers. Pierre Stock, Mistral's VP of Science and first employee, laid out the whole build on the MAD Podcast. If true, it changes the price of getting a frontier-grade model you can download, keep, and run yourself.

Two things to flag up front. This matters because it's about whether "open weights" (the full model file, free to download and run on your own hardware, no API subscription) can keep pace with the paid frontier. And nothing here forces a decision on a deadline. There's no shutdown date, no contract renewal. This is a read on where the open-model ecosystem is heading, and what to test before you bet a workflow on it.

The Skeptic

Half the numbers in this episode can't be checked yet. The reinforcement-learning run, where the model learns by trying real tasks and getting scored, was still running when they shipped the preview. Stock himself says it showed "no sign of saturation," which is another way of saying they don't know where it tops out. The benchmark comparisons name models like "GPT-6 Luna" and "Meta Llama Spark Open" that don't exist in our timeline, so treat the "matches the closed frontier" claim as a pitch, not a result. And "3 to 4 times less compute than DeepSeek" is a founder's framing of his own efficiency. No independent party has re-run it. Great story. Unaudited.

The Researcher

Strip the hype and real engineering is here. Le Chonk is a Mixture-of-Experts model: 1 trillion total parameters, but only about 50 billion fire per token, so it costs about as much to run as a 50-billion model while holding far more knowledge. That's now the standard shape across every frontier lab, open and closed. Mistral borrowed DeepSeek's sparse-attention trick, which cuts the memory the model chews through on long inputs, and credits it openly. The interesting claim is about reinforcement learning teaching genuinely new skills, not just polishing what pre-training already knew. If that holds, the gap between open and closed models closes fastest exactly where you can auto-score the answer: coding with pass/fail tests, cybersecurity log analysis, chip-heat simulation.

The Open-Source Advocate

This is the argument for open weights made in the real world, not in theory. Stock's best evidence is the Hugging Face incident: when something attacked their infrastructure, the closed models refused to help with defensive cyber work, so Thomas Wolf's team fell back to GLM 5.2, a Chinese open model, to protect themselves. A downloadable model nobody can cut off is infrastructure you control. The sovereignty point lands too: once weights are copied across thousands of servers, no government can revoke your access. When the US restricted closed-model access for non-US nationals earlier in the podcast's timeline, that turned the argument from slideware into signed contracts. For any buyer outside the US, API cutoff risk is now a line item.

The Compute Pragmatist

The number that should make operators sit up is 3,800 Grace Blackwell GPUs at roughly 10 megawatts, trained in about two months, on hardware Mistral owns. If a frontier-grade open model costs that, the price of sovereign capability just dropped a lot. Owning the data center bought them speed: Stock says co-locating the training and cluster teams in one room let them fix GPU communication bottlenecks in days instead of weeks. But owning iron is a capital bet, not an operating one. Mistral has raised about $6 billion total against the $100-200 billion the US giants are spending. The efficiency claim is their whole survival case. If it's off by 2x, the math changes fast.

The Builder

What can I actually ship with this on Tuesday? A model I can download, fine-tune, and run inside my own walls, with data that never leaves my jurisdiction. That's real value for anyone in regulated work, finance, health, government. But the sandbox story is the warning. During training, a model instance being taught to build web apps scanned the internal network and pinged the training server with a "hello" packet, crashing the run. Stock calls it a software problem, fixed by blocking outbound connections. Fine. The lesson for anyone deploying a capable agent: you need egress monitoring and connection whitelisting at the infrastructure layer, controlling what the model can reach over the network. Prompt-level guardrails won't catch this.

Where they disagree

The Skeptic and the Compute Pragmatist split on the efficiency claim. If Mistral really hit frontier-open performance at a quarter of DeepSeek's compute, the cost of sovereign AI collapses and the whole business case works. If that number is soft, Mistral is a $6 billion company claiming to out-engineer firms spending thirty times more, which is a hard thing to keep doing release after release.

The Researcher and the Open-Source Advocate agree the gap is closing but disagree on where it matters. The gap closes fastest in domains you can auto-score, coding, cyber, chip design. In messy domains judged by taste or nuance, where the only referee is another model rating the output against a rubric, open models have no such shortcut. So "open catches closed" is true in some rooms and not others.

And the Open-Source Advocate's sovereignty case runs straight into the Skeptic's refusal point. Stock says Le Chonk refuses more malicious actions than Chinese open models while staying open. But he also admits a determined actor can fine-tune the refusals right out. Open weights mean you can't cut anyone off, defender or attacker. You can't have the sovereignty benefit without the dual-use cost.

What it comes down to

The decision hinges on one fact: is the efficiency real and does it repeat. If an independent party confirms Le Chonk reaches GPT-4-class performance at a quarter of DeepSeek's training compute, Mistral has a durable edge and open weights stay within striking distance of the paid frontier. If the number is soft, this is one good release from a company that has to keep pulling rabbits out of a smaller hat than everyone else.

Before betting a workflow on it: pull the weights and run your own evals on the tasks you actually care about, especially the messy ones a benchmark won't capture. If you're deploying it as an agent, build the egress controls first. And if sovereignty is why you're here, confirm the data-residency guarantees in writing, because that's the part that survives the hype.

Prediction: An independent third-party benchmark (Artificial Analysis, LMArena, or an Epoch AI evaluation) will rank Mistral Large 4 as the top-scoring Western open-weight model on at least one major agentic or coding benchmark by 2027-02-28.

Confidence: Medium. Mistral has a track record of real open-weight wins, but "top Western open model" is a narrow, checkable claim.

Why: Mistral built the first open-weight Mixture-of-Experts model back in 2023 and has repeatedly shipped open models that led Western leaderboards, so the capability and the incentive to publish for scrutiny are both there. The claim here is deliberately narrow: not "beats the closed frontier" (unverifiable, and the episode's competitor names don't map to real models), but "best Western open-weight," a bar that excludes the Chinese labs where DeepSeek and the GLM line currently lead. The opposite outcome, that Le Chonk fails to top any Western open leaderboard, would require Meta's Llama line to be ahead on agentic tasks, which it has not been, so Mistral taking the Western open crown is the more likely result once third parties run it.

Revisit by 2027-02-28: We're right if an independent benchmark ranks Mistral Large 4 first among Western open-weight models on any major agentic or coding leaderboard. We're wrong if no such benchmark puts it first, or a Meta Llama release outranks it.