The Mistral Large / Medium / Mixtral family is the frontier model line developed by Mistral AI, the only Europe-headquartered frontier-class AI lab. The line began with Apache 2.0 open-weight releases in 2023 and shifted toward closed and research-licensed flagships from 2024 onward, making it a commonly cited reference case in debates over open weights and EU sovereign AI.
| Field | Value | |
|---|---|---|
| Developer | [[mistral-ai | Mistral AI]] (Paris, France) |
| Founded | April 2023 | |
| Key people | Arthur Mensch (CEO), Guillaume Lample, Timothée Lacroix | |
| Lead flagship | Mistral Large 3 (Nov 2025), closed API | |
| Open weights | Frontier models closed; small models under Apache 2.0 / Mistral Research License |
Mistral AI is the only Europe-headquartered frontier AI lab and an explicit focus of EU sovereign-AI policy.
Model lineage
| Model | Release | Architecture | License | Notes |
|---|---|---|---|---|
| Mistral 7B | Sep 2023 | Dense 7B | Apache 2.0 | Sliding-window attention; instant open-weight standard |
| Mixtral 8x7B | Dec 2023 | MoE — 46.7B total / 12.9B active (8 experts / 2 active) | Apache 2.0 | First widely deployed open MoE |
| Mistral Large | Feb 2024 | Closed dense ~123B (unconfirmed) | Commercial API only | First Mistral closed-weights flagship |
| Mixtral 8x22B | Apr 2024 | MoE — 141B total / 39B active | Apache 2.0 | Last open flagship-scale release |
| Mistral Large 2 | Jul 2024 | Dense 123B | Mistral Research License (non-commercial) | Partial opening — weights available, commercial use requires license |
| Mistral Nemo | Jul 2024 | Dense 12B (w/ NVIDIA) | Apache 2.0 | Multilingual workhorse |
| Codestral | May 2024 | Dense 22B | Mistral Non-Production License | Code-specialized |
| Mistral Medium 3 | May 2025 | Closed (details withheld) | API only | Positioned as enterprise default |
| Mistral Large 3 | Nov 2025 | Closed (details withheld) | API only |
Capabilities and benchmarks
At its February 2024 release, Mistral Large was marketed as the second-strongest non-US model on MMLU (81.2%), behind GPT-4 (86.4%) and ahead of Claude 2.1 and Gemini Pro. By late 2024, Mistral Large 2 (MMLU 84.0%, HumanEval 92.0%) approached parity with Llama 3.1 405B at roughly a quarter the size. Benchmark parity with frontier US closed models narrowed rather than closed across 2025.
Training and architecture
The line uses sliding-window attention (introduced with Mistral 7B) and grouped-query attention across subsequent models. The Mixtral models are mixture-of-experts (MoE): Mixtral 8x7B has 8 experts with 2 activated per token via a learned router, the first widely deployed non-Chinese open MoE architecture, totaling 46.7B parameters with 12.9B active; Mixtral 8x22B totals 141B parameters with 39B active. Mistral Nemo introduced the Tekken tokenizer, which improved multilingual handling.
Mistral has not published a full training-compute disclosure for its flagship models. Mistral Large is rumored to have trained on fewer than 5,000 H100-equivalents, small by frontier standards, though this is not confirmed.
Licensing trajectory
The family's movement from open-default to closed-flagship is a commonly cited counterpoint to the expectation that European AI would remain open.
From September 2023 through April 2024, Mistral released models under Apache 2.0 by default; Mixtral 8x22B was at the time the largest openly licensed MoE. In February 2024, Mistral Large was released closed-weights simultaneously with a Microsoft partnership (a €15M investment and Azure distribution), prompting the first wave of criticism that Mistral was abandoning open source. In July 2024, Mistral Large 2 was released under the Mistral Research License, with weights inspectable but commercial use gated; Mistral framed this as "open weights, protected commercial rights." Across 2025, Medium 3, Large 3, and subsequent flagships were closed entirely, while smaller models (Mistral Small 3, Nemo, Ministral) continued under Apache 2.0 or the Mistral Research License.
As of April 2026 the pattern held: frontier Mistral models are closed, mid-tier models are research-licensed, and small models remain open. Open-Source AI situates this within the broader range of openness terms that have emerged across labs.
Geopolitical role
Mistral is a frequently cited case study in EU sovereign AI. French President Emmanuel Macron has repeatedly named it as strategic infrastructure, and it has received French and EU industrial-policy support and figures in the "EU AI Champions" discourse. Mistral was a core participant in EU AI Act negotiations and lobbied successfully for lighter general-purpose AI (GPAI) obligations on open-weight models; its lobbying is central to the GPAI code-of-practice carve-outs for open models in the EU AI Act. The February 2024 Microsoft partnership drew simultaneous criticism from French open-source advocates and EU industrial-sovereignty voices.
In AI sovereignty debates, Mistral is the primary exhibit for the argument that "European AI" is a coherent industrial-policy category, and in geopolitical framings such as China–US AI race discussions it is the most common center of the third-pole narrative positioning Europe as neither US nor China.