Mistral debuts Large 4 ‘Le Chonk', a 1 Trillion parameter text output model with high benchmarks planned for open weights release

Jessica One trillion parameters is the flashy part. The real thing here is Mistral saying, pretty plainly, that the weights are coming on October twenty-seventh, not some vague someday. That’s the bit people can actually plan around.

Cathy Yeah, and that’s the only part I trust yet. Preview numbers are cheap, especially when the company is still doing reinforcement learning on the final checkpoint.

Jessica Right, but the rollout itself is the product. If you’re a security team or a government shop, a model you can eventually run yourself is a different sale than a nice demo in a browser.

Cathy Mm-hm.

Jessica And Mistral is being very Mistral about it. Public preview now, weights later, custom license, enterprise control, sovereign infrastructure, zero-data-retention options. That’s not just model theater. That’s a deployment story.

Cathy Sure, but the deployment story only works if the released thing is actually usable. One trillion total parameters sounds huge, but only forty-nine billion are active at inference, so the practical question is cost per token and whether the routing behaves.

Jessica Exactly, and that sparse piece is what makes it interesting. It’s the same general move they’ve been making for a while, just at a much larger scale. Also, Le Chonk is a ridiculous name, which I respect deeply on this podcast called Exploring Next.

Cathy Stop it.

Jessica No, come on, you know the meme helped. Le Chaton Fat was already living rent-free online, so Mistral basically walked in and said, fine, here’s the actual cat.

Cathy That part is kind of funny, annoyingly. But the benchmark picture is where I get prickly, because the charts are doing a lot of work for them.

Jessica You mean the coding chart?

Cathy Yeah. ML4 at sixty-two percent on DeepSWE looks solid, especially against the other open-weight systems they listed. But the live leaderboard has stronger published configurations for models like GLM-5.3 and Kimi K3, so this is competitive, not a clean coronation.

Jessica Which is still a meaningful ship moment. For the user, the question is not, did it win every public chart. It’s, can I get a frontier-ish open model that’s good enough for coding, cyber, finance, and weird document-heavy workflows without begging a closed provider for permission?

Cathy Right, and cybersecurity is the most interesting target there. If you’re doing defensive testing or code scanning at scale, an open-weight model with fewer moderation surprises can be a real operational win.

Jessica I buy that. The enterprise story is basically the boring-but-works tattoo again: control layer, receipts, deployment you can own, the whole thing. That’s where adoption happens, not in the benchmark victory lap.

Cathy Mm-hm. And the legal benchmark is the cleaner signal anyway. Their fifteen percent on Harvey’s Legal Agent Benchmark is at least in the same universe as the public comparables, which makes the claim feel less like slideware.

Jessica That’s the part I’d put money on getting attention fast. I think by the end of the month, at least one serious public evaluation will put ML4 in the top tier of open-weight models on a real workflow benchmark, not just a pretty chart.

Cathy I’d go lower than that. Sixty-forty the released weights land and outsiders reproduce something close to the preview results within two weeks of October twenty-seventh. If that holds, then we’ll know this was more than a launch-day flex.

Jessica That’s a very you bet, and annoyingly fair. Also, we’re almost a year into doing this show and somehow still arguing about whether the cat model is real enough.

Cathy Amazingly, yes.

Jessica Okay, when the weights drop, send me the first repo people use. I want the actual running story, not the marketing version.

Cathy Deal. And if the license turns into a mess, I’m blaming the cat.

Jessica Fair. Come back when it’s furry and reproducible, Cathy.