16 models · on the curve since 2023

Mistral AI

Value position · the site’s read Output layer

Mistral ships open weights, the fastest-commoditising part of the stack. The durable position is European sovereignty, on-prem enterprise trust, and Le Chat as a distribution surface.

What would change this readOpen weights stop tracking the closed frontier, or sovereignty stops being something European buyers will pay for.

Recent cadence
6 in last 12 mo
4 the year before
Landmark density
4 / 16
defining releases
Weights
Open + closed
Mistral Large 3 flagship

Founded 2023Paris, France

The substrate underneath figures as of Aug 2026

Trains and serves on

NVIDIA throughout. Mistral Large 3 was trained from scratch on 3,000 NVIDIA H200 GPUs, and the whole Mistral 3 family on Hopper silicon. Mistral Compute, its own sovereign cloud, runs GB200/GB300/B300 accelerators and targets 200 MW of EU capacity by 2027; a July 2026 Microsoft partnership adds thousands of Vera Rubin GPUs.

ConcentrationEvery disclosed training run and every announced buildout uses NVIDIA silicon, so the European sovereign position is sovereign in ownership, siting and jurisdiction — but not in chip supply.

Flagship, per million tokens

$0.50 in $1.50 out

Mistral Large 3, its open-weight flagship

Was

$2 / $6

Mistral Large, legacy tier · 2024

Weights
Mostly open. Mistral Large 3, Ministral 3, Small 4, Voxtral and Leanstral under Apache 2.0; Medium 3.5 under a modified MIT with a large-revenue carve-out; OCR, Codestral and the embedding models stay closed.
Where you buy it
La Plateforme and Le Chat · HuggingFace · Amazon Bedrock · Microsoft Foundry and Azure · IBM watsonx
Structure
Mistral AI SAS, a French société par actions simplifiée. Private, unlisted, founded April 2023 in Paris. No dual-class or foundation structure is documented.
Raised
About $4B equity plus $830M debt raised in March 2026 for a Paris-region datacentre
Valuation
EUR 11.7B post-money (Series C, Sep 2025, led by ASML)
Revenue
ARR above $400M (Feb 2026, attributed to the CEO), up from about $20M a year earlier

Biggest single dependencyThe entire disclosed compute base, built and planned, is NVIDIA silicon procured through partners including Microsoft and NVIDIA, both of which are also investors.

Flagship is genuinely ambiguous here: Mistral labels Large 3 its flagship, yet Medium 3.5 is listed as the frontier-class model and costs three times as much on input and five times on output. A round reported at about EUR 20B has not closed and is excluded; the verifiable valuation is the Sep 2025 one. Revenue is a CEO statement, not a filing. Headcount is not published and trackers span 350 to 1,392, so it is omitted. An open-weight frontier model was trailed for summer 2026 and had not shipped as of 1 Aug.

On the durability curve frontier era
2023202420252026 Mistral 7BRobostral Navigate ARCHITECTURE the substrate · compounds  → OUTPUT per model · commoditises  → CROSSOVER the durability thesis  → DURABLE VALUE-CAPTURE ↑ the margin a release keeps · not raw capability TIME · FRONTIER MODEL RELEASES →
Output · value a model keeps Architecture · the compounding substrate Mistral AI release

Hover any point to read the release. The y-axis is qualitative; only the x-axis carries dates.

When to reach for them practical read
Open weights
Reach for it when you need to self-host or fine-tune freely (Mistral 7B, Sep 2023).
On-prem & sovereignty
Pick Mistral where data residency is a requirement (Mistral Large, Feb 2024).
Efficient open MoE
Use Mixtral-class models for strong quality per cost (Mixtral 8x7B, Dec 2023).
The landmark releases 4 of 16 models
  1. Dec 2025 Mistral Large 3 Mixture-of-experts flagship MoEFlagship
  2. Jun 2025 Magistral First reasoning models ReasoningFirst of kind
  3. Dec 2023 Mixtral 8x7B Open mixture-of-experts breakthrough OpenMoE
  4. Sep 2023 Mistral 7B First model, punched far above its size OpenFirst model
Full release logall 16, newest first
  1. 2026
  2. Jul Robostral Navigate First embodied-navigation model Robotics
  3. Jul Leanstral 1.5 Formal proof and verified coding Proofs
  4. Jun Mistral OCR 4 Document intelligence Documents
  5. May Mistral Medium 3.5 Mid tier, powering remote coding agents Medium tier
  6. Mar Mistral Small 4 Small tier refresh OpenSmall tier
  7. 2025
  8. Dec Mistral Large 3 ◆ Mixture-of-experts flagship MoEFlagship
  9. Jun Magistral ◆ First reasoning models ReasoningFirst of kind
  10. May Mistral Medium 3 Mid-tier flagship Mid-tier
  11. Jan Mistral Small 3 Efficient open model OpenEfficient
  12. 2024
  13. Sep Pixtral First multimodal model OpenMultimodal
  14. Jul Mistral Large 2 123B-parameter flagship 123BFlagship
  15. May Codestral 22B Dedicated code model OpenCoding
  16. Apr Mixtral 8x22B Larger open MoE OpenMoE
  17. Feb Mistral Large First proprietary flagship ClosedFlagship
  18. 2023
  19. Dec Mixtral 8x7B ◆ Open mixture-of-experts breakthrough OpenMoE
  20. Sep Mistral 7B ◆ First model, punched far above its size OpenFirst model
Where they sit on weights versus the field
Open weights
DeepSeek Moonshot AI Zhipu (Z.ai)
Mixed
Google Meta Mistral AI Alibaba (Qwen)
Closed / API
OpenAI Anthropic xAI
The durability take on Mistral AI from the archive
The Model Is Not the Moat If frontier capability keeps centralising, the durable edge shifts outward into trust, workflow fit, and the surrounding package. Read → The Forest Floor Is The Product Most people are optimising for canopy. The work that survives the next five AI releases is built from the forest floor up. Read → You're Not Comparing Models. You're Comparing Contracts. Agent benchmarks don't measure models. They measure contracts. Two teams running the same model can publish different scores, and both can be honest. Read →
About Mistral AI at a glance
Founded
2023
HQ
Paris, France
Leadership
Arthur Mensch (CEO)
Founders
Mensch, Lample, Lacroix
Flagship
Mistral Large 3
Weights
Open + proprietary
Also builds
Le Chat, La Plateforme
Backers
a16z, General Catalyst

Founded April 2023 by Arthur Mensch, Guillaume Lample and Timothée Lacroix, former DeepMind and Meta researchers. Headquartered in Paris, it is Europe’s leading frontier lab, known for high-quality open-weight models alongside proprietary ones and its assistant, Le Chat.

The frontier chart → Read the archive →