Mistral AI

Mistral AI is a Paris company that trains open-weight large language models, and it is the only European lab in this section. Its distinguishing fact today is one that would have surprised anyone tracking it two years ago: its most capable model is Apache 2.0. Mistral Large 3 — 675 billion total parameters, 41 billion active — sits in Mistral’s documentation under open weight models, not under the premier tier.

The company was founded in April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix. The provenance matters: Lample and Lacroix were authors on Meta’s LLaMA paper and Mensch on Google DeepMind’s Chinchilla scaling paper, so the founders wrote two of the papers the rest of this section rests on before leaving to compete with their former employers.

Two models that taught the field something#

Mistral 7B (September 2023) was the model that made “small and good” a serious position. It combined grouped-query attention with sliding-window attention — each layer attending only to a fixed span of recent tokens rather than the whole sequence — and demonstrated that careful design beat raw parameter count at that scale. For a great many people it was the first capable model they ran on their own hardware.

Mixtral 8x7B (December 2023) was the first widely-used open sparse mixture of experts, and its name is the single most instructive piece of arithmetic in this section. 8x7B reads as eight 7-billion-parameter models stapled together — 56 billion — but the model holds 46.7 billion parameters and uses 12.9 billion per token, because only the feed-forward block is replicated and everything else is shared once. Its own page works the numbers through.

The licence split#

Mistral uses four sets of terms, not two, and the difference between them decides whether a model can be deployed at all:

Alongside these sit genuinely closed models — the embedding, moderation, document-recognition and speech lines are interface-only.

The practical consequence is that “is it a Mistral open model?” is not a question with one answer. Two models released months apart can differ on whether a business may run them at all, and the model’s name does not tell you which.

The reversal, and the criticism that is now out of date#

The standard criticism of Mistral is that it built its reputation on Apache 2.0 releases and then reserved its best work for a paid interface. For 2024 and 2025 that was accurate: Mistral Large was interface-only, Mistral Large 2 shipped weights under a research-only licence, and Codestral got a bespoke non-production licence.

What has happened since is a reversal. Mistral Large 3, released December 2025, is Apache 2.0. Mistral Medium 3.5, the most expensive model in the lineup, publishes its weights under Modified MIT. The company’s funding announcement is titled “Making sovereign, open-weight AI the technology frontier,” and argues that open weights are what prevent customers being “locked into a single vendor’s roadmap, pricing or availability.”

Both halves are the story: a company that drew its early reputation from open releases, drifted away from them, and has since come back further than it started.

The sovereignty argument#

Mistral’s positioning is not mainly about capability rankings, and treating it as a smaller American lab misses what it sells. Its stated target customers are enterprise and government in finance, manufacturing, defence, energy and public administration — buyers for whom where the model runs and who controls it is a procurement requirement rather than a preference.

The funding follows that positioning rather than the usual venture pattern: a Series C led by the Dutch semiconductor-equipment maker ASML in September 2025, then a Series D of about €3 billion led by Samsung Electronics, with the Grand Duchy of Luxembourg among the investors. Strategic industrial and sovereign money, not growth capital.

This is where open weights stop being an ideological position and become a product feature. A European bank or defence ministry that cannot send data to a United States interface, and cannot accept that a vendor might withdraw or reprice a model mid-contract, has a short list of options. Mistral is on it; most of this section is not. The same logic explains why Google markets Gemma for air-gapped deployment.

Consolidation#

Mistral once ran parallel specialised lines — Magistral for reasoning, Devstral for code, Pixtral for vision. Those are retired, all on or before 31 July 2026, replaced by Mistral Medium 3.5 and Mistral Small 4. Mistral Small 4 “consolidates three specialized models into one” with a reasoning_effort setting, which is the same consolidation OpenAI made when GPT-5 absorbed the o-series and DeepSeek made when the R line merged back into V. Three labs of very different sizes reached the same conclusion within about a year: a reasoning model is better as a setting than as a separate product.

Status#

Checked 11 September 2026. Mistral’s model overview is the live source, and its licence explainer is the authority on which terms apply to which model.

Mistral Large 3 was released 2 December 2025, Mistral Small 4 on 16 March 2026 and Mistral Medium 3.5 on 28 April 2026. Closed models cover embeddings, moderation, document recognition and speech.

Sources#

• Jiang et al., Mistral 7B, arXiv:2310.06825 (2023)
• Jiang et al., Mixtral of Experts, arXiv:2401.04088 (2024)
• Mistral AI, Mixtral of experts (11 December 2023) — the 46.7B / 12.9B figures
• Mistral AI, Mistral 3 (2 December 2025)
• Mistral AI, Mistral Small 4 (16 March 2026)
• Mistral AI, Under which license are Mistral’s open models available?
• Mistral AI, Making sovereign, open-weight AI the technology frontier
• Mistral AI, About — founding and positioning