Run Mistral locally: Europe's open family, and the license lines you can't cross

A developer types ollama run codestral, and within minutes has one of the best open code models in the world running on a laptop — completing functions, explaining stack traces, refactoring on command. It is fast, it is good, and it is free to download. So they wire it into the product they’re shipping.

That last step breaks the license. Codestral, the Mistral model developers most often reach for, ships under a non-production license— research and testing only. Using it in a real product is exactly the thing the terms forbid. This is the Mistral paradox in one move: a family that is mostly, genuinely open source still hides a few models behind lines you can’t cross — and the model you want most is sometimes the one you’re not allowed to use. The fix isn’t to avoid Mistral. It’s to read the map.

Mistral is the open-model family that comes with fine print

Mistral AI is a Paris lab founded in 2023 by alumni of Google DeepMind and Meta. It made its name by giving things away: its first model, Mistral 7B, landed in September 2023 under an Apache 2.0 license and claimed to “outperform Llama 2 13B on all benchmarks” at roughly half the size. For a year, “Mistral” was shorthand for “the good open model you can actually run.”

Then the catalog grew, and the licensing forked. Some models stayed fully open. Others moved behind a research license or a non-production license — free to tinker with, off-limits to ship. A reader who assumes “Mistral = open” can get burned, and the burn is legal, not technical. The model runs fine; it’s the lawyer who says no. So before any roster or run command, the useful unit of this family is the license column.

Europe’s only frontier lab, and that’s the pitch

There is one fact about Mistral that no US or Chinese lab can copy: it is European. Mistral is the only company building frontier-scale models inside the EU, and it has leaned into that as a business. Its API, La Plateforme, runs on infrastructure hosted exclusively in the European Union, with a GDPR-compliant data-processing agreement available by default. For a German hospital or a French bank, “the model never leaves the EU” is not a nice-to-have — it’s the thing that lets the project happen at all.

The market is rewarding the position. Mistral’s annual recurring revenue ran from about $16M at the end of 2024 to roughly $400M by January 2026 — a roughly 20x jump in a year, built largely on regulated European enterprises that need a model and a compliance story in the same vendor. If you care where your data sits, this is the family whose entire identity is the answer. It pairs naturally with the case for keeping regulated AI work off the open cloud.

The advantage is structural, not marketing. As an EU company, Mistral falls directly under the EU AI Act and European data-protection law, rather than promising compliance from outside the bloc. For a buyer in a regulated sector, that distinction is the difference between a model their legal team can sign off on and one that triggers a cross-border transfer review. No US or Chinese lab can offer the same posture, however good its model — which is why Mistral’s pitch lands hardest exactly where the data is most sensitive.

The open tier is the real thing — Apache 2.0, no asterisks

Start with the half that’s genuinely open, because it’s most of the family. Apache 2.0 is the same license that covers thousands of everyday software libraries. It lets you run the model commercially, fine-tune it, serve it to paying customers, and redistribute it — no royalty, no revenue share, and crucially no user cap. That last point is the contrast with Llama’s Community License, which adds a 700-million-user ceiling and an attribution rule. With Apache, there is no ceiling and no string.

The open tier’s standout contribution is the mixture-of-experts model. When Mistral released Mixtral 8x7B in December 2023, and later Mixtral 8x22B — 141 billion total parameters, but only 39 billion firing on any given token — it made open MoE mainstream. The trick is efficiency: you get the quality of a huge model while only paying compute for a slice of it on each word. Most of today’s efficient open models, Mistral’s own flagship included, descend from that idea.

Efficiency is the family’s signature, and it’s why these models run well on hardware you already own. Mistral 7B’s launch claim wasn’t just that it beat Llama 2 13B — it was that it performed like a Llama model “more than 3x its size” on reasoning and STEM. That obsession with quality-per-parameter carries through the whole open tier: small models that punch above their weight, plus open specialists like Mistral NeMo and the math-tuned Mathstral. It’s the opposite of the “bigger is always better” race, and it’s what makes the Apache lineup a serious option for a laptop instead of a data center.

The open tier also covers the specialists. Pixtral 12B reads images and screenshots under Apache 2.0. Devstral and Codestral Mamba handle code, openly. And Mistral Small — a 24B model that fits a 16 GB laptop once quantized — is the everyday workhorse, the one most people should install first. If you only remember one rule: the Apache models are yours to keep, forever, on a machine you own.

The lines you can’t cross: Codestral and the older flagships

Now the amber rows. Three kinds of restriction show up in the Mistral catalog, and they sting in different ways. The sharpest is Codestral 22B, the popular code-completion model. It ships under the Mistral Non-Production License: you may use it for research and testing, but a commercial deployment requires contacting Mistral for a paid license. The model is a 13 GB download that runs with a single ollama run codestral — which is exactly what makes it a trap. Nothing stops you technically; the terms stop you legally.

The second kind is the research license. The previous flagship, Mistral Large 2, and the big vision model, Pixtral Large, both ship under the Mistral Research License — fine for experiments and evaluation, not for a product. The third kind is older still: the original Ministral edge models from October 2024 were commercial- or research-only — the 3B you couldn’t even self-host freely.

The practical move is substitution. Want local code you can ship? Use Devstral or Codestral Mamba, both Apache, instead of Codestral 22B. Want a big open flagship? Reach for Mistral Large 3, not Large 2. The restricted models are usually a free, open sibling away — you just have to know which door says “research only.”

Mistral 3 swung the whole family open

The split is narrower than it used to be, because Mistral’s biggest release bet on openness. On December 2, 2025, the lab shipped Mistral 3 — and put the entire generation under Apache 2.0. That includes Mistral Large 3, a sparse mixture-of-experts with 41 billion active and 675 billion total parameters that debuted at #2 among open non-reasoning models on the LMArena leaderboard. A frontier-class flagship, fully open, was not the norm a year earlier.

The same release relaunched the edge tier as Ministral 3— dense 3B, 8B, and 14B models with native vision, each now Apache 2.0. Note the arc: the original 2024 Ministral was the restricted one you couldn’t self-host; its 2025 successor is free to run and sell. The 14B reasoning variant hits 85% on the AIME ’25 math exam, and the family is built for exactly the on-device jobs — translation, voice assistants, local analytics — that belong on your own hardware, not a rented server.

One new wrinkle is worth knowing, because almost nobody reads it. A few recent models use a modified MIT license that grants the same freedoms as Apache, with a single exception: companies with monthly revenue above $20M USD must either buy a commercial license or use Mistral’s hosted Studio. For a solo developer or a small team, this changes nothing — you are nowhere near the threshold. For a large enterprise, it’s a line item to check before deployment. Either way, it’s a reminder that “open” in this family always rewards a second look at the terms.

Pick by your hardware, then run one command

With the legal map clear, the practical question is which open model your machine can hold. The download size is a fair proxy for the memory you’ll need, and the rule from the broader local-models guide applies: install Ollama, pick the largest model that fits, and leave headroom for context.

For most readers the answer is the middle of that table: ollama run mistral-small gives you a capable 24B assistant that reads images and fits a typical 16 GB laptop, entirely offline. Drop to ollama run mistral (the 7B) on an 8 GB machine, or down to the Ministral 3 edge tier on a phone or single-board computer. Step up to Mixtral if you have the RAM and want the mixture-of-experts headroom. Every command in that table lands on an Apache model — nothing in it carries the production restriction.

Prefer not to run anything locally? The same models are cheap on La Plateforme — Mistral Large 3 bills around $0.50 per million input tokens and $1.50 per million output, and Ministral 3 8B runs about $0.15 per million each way — with the EU hosting baked in. But the whole appeal of the open tier is that you don’t need the API at all. This is the same own-it-versus-rent-it choice that runs through every local-model decision.

So which Mistral should you run?

For a general assistant on the machine you already own, install Mistral Small (24B) and stop overthinking it. For edge and on-device work, the Ministral 3 tier is purpose-built and Apache-licensed. For local code you intend to ship, reach for Devstral or Codestral Mamba — and treat the headline Codestralas a research toy unless you’ve bought the commercial license. For anything touching EU-regulated data, La Plateforme’s EU hosting is the family’s real edge.

The larger lesson outlives any single model name. Mistral proved that “open source” in AI is not one thing but a spectrum — Apache on one end, research-only on the other, a $20M tripwire in between. With Mistral 3 the family swung hard toward the open end, which is good news for everyone with a laptop. But the half-open reputation is earned, and the only protection is the habit of checking the license column before you build. Read the map, and almost all of Europe’s best open models are one command away.