AI Policy Wiki
Dashboard

Mistral Large / Mistral Medium / Mixtral

high confidence · updated 2026-06-06

Mistral AI's frontier family — Mixtral 8x7B (Dec 2023), Mixtral 8x22B (Apr 2024), Mistral Large (Feb 2024 → Large 2 Jul 2024), Mistral Medium 3 (May 2025). Europe's only frontier-class model family and the reference case for a partial retreat from open weights.

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.

FieldValue
Developer[[mistral-aiMistral AI]] (Paris, France)
FoundedApril 2023
Key peopleArthur Mensch (CEO), Guillaume Lample, Timothée Lacroix
Lead flagshipMistral Large 3 (Nov 2025), closed API
Open weightsFrontier 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

ModelReleaseArchitectureLicenseNotes
Mistral 7BSep 2023Dense 7BApache 2.0Sliding-window attention; instant open-weight standard
Mixtral 8x7BDec 2023MoE — 46.7B total / 12.9B active (8 experts / 2 active)Apache 2.0First widely deployed open MoE
Mistral LargeFeb 2024Closed dense ~123B (unconfirmed)Commercial API onlyFirst Mistral closed-weights flagship
Mixtral 8x22BApr 2024MoE — 141B total / 39B activeApache 2.0Last open flagship-scale release
Mistral Large 2Jul 2024Dense 123BMistral Research License (non-commercial)Partial opening — weights available, commercial use requires license
Mistral NemoJul 2024Dense 12B (w/ NVIDIA)Apache 2.0Multilingual workhorse
CodestralMay 2024Dense 22BMistral Non-Production LicenseCode-specialized
Mistral Medium 3May 2025Closed (details withheld)API onlyPositioned as enterprise default
Mistral Large 3Nov 2025Closed (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.

Relationships