Mistral put Large 4 out as an API-only preview on October 6, and in late October it released the weights. The internal nickname is “Le Chonk,” which is fair: 1.05 trillion parameters in total, with 52 billion active for any given token.

Back at the preview, TechCrunch reported that Mistral planned to open the weights roughly three weeks after safety testing, and that’s more or less how it played out. What’s on Hugging Face now isn’t just the preview with the lock removed, though. It’s a newer build with more reinforcement learning, and a few things moved in ways that matter.

The real improvement is fewer tokens

By Mistral’s numbers, Terminal-Bench 4.0 went from 28% in the preview to 41%, while output tokens for the same work dropped 30%. On Artificial Analysis’s Coding Agent Index it lands at 48%, about level with GLM-5.3, but uses 47% fewer tokens per task on average.

Almost all of that is self-reported, and some of it comes from private benchmarks nobody else can rerun. So I wouldn’t lean on the leaderboard spots. The direction is more telling: this update is less about getting smarter and more about getting to the same answer in fewer tokens. If you’re paying per token, that’s the number that shows up on your invoice.

Watch the default reasoning mode

There are three reasoning levels: Low, Medium and High. The default is High. Leave it alone and even trivial requests run in the mode that thinks the longest.

Pricing is $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. Output costs three times as much as input, so long reasoning traces add up fast. If you’re wiring this into classification, extraction or quick summaries, set it to Low first. Mistral says Low replaces the old “no reasoning” option, so that’s what it’s meant for.

The flip side: workloads that resend the same long context are where this model looks good. You get a 1M-token context window, and cached input is a tenth of the normal input price.

Open weight, with one commercial string attached

The license is the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. You can use it, modify it, fine-tune it and build commercial products on it. The exception is selling the model itself as a hosted service. A Model-as-a-Service business needs a separate commercial license from Mistral.

Running it internally for document processing or coding help is fine. Spinning up an inference API and reselling access means talking to Mistral first. If you see it described as “fully open source, no strings,” that detail got lost.

Can you run it yourself?

Mistral doesn’t list hardware requirements, so here’s the back-of-the-envelope version. At 4-bit, each parameter takes half a byte, which puts the weights alone at about 525GB. At 8-bit, double that. Then you still need memory for the context.

Because it’s a mixture-of-experts model, each token only touches about 52B parameters, so it’s quick for its size. But every weight still has to be loaded. In practice that means a multi-GPU server, not a home rig. If you’re into local models, keep an eye out for smaller derived or distilled versions instead.

Why European companies should care

The clearest winners are organizations with strict data-residency rules. If sending data to a US or Chinese API is a non-starter, being able to self-host a frontier-class model from a European lab is a big deal on its own. Mistral leaned into exactly that framing with TechCrunch, pitching itself as an alternative to both closed models and the open models coming out of China.

The training claim is worth a mention too. Pierre Stock, Mistral’s VP of Science, said the model was trained on roughly 4,000 Nvidia GPUs, which he put at two to three times fewer than Chinese competitors use. There’s no way to verify that from the outside. If it holds up, though, it chips at the assumption that frontier models need the biggest compute budgets around.

Still unconfirmed

Independent benchmarks aren’t in yet. Security deserves attention too: Mistral says it can’t fully rule out the model giving real uplift to attackers and is adjusting its guardrails. A security-focused variant is only open as a preview to vetted partners such as Orange Cyberdefense.

Once third-party numbers land, I’ll check whether the preview-to-release jump really looks this big.