HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Mistral AI—Mistral ComputeMistral AI—Mistral ComputeMistral AI—Mistral modelsMistral AI—

Mistral models

by Mistral AIFR

ScalingStage 5 of 5

Mistral 3 open-weight family released December 2025, plus paid and reasoning models.

Updated 2 Dec 2025Checked 25 Sep0 updates this week

Milestones

No announced next step
  1. Mistral 7B released with open weights27 Sep 2023Complete.
  2. Mixtral 8x7B mixture-of-experts modelDec 2023Complete.
  3. Mistral Large 224 Jul 2024Complete.
  4. Magistral reasoning models10 Jun 2025Complete.
  5. Mistral 3 family including Mistral Large 3Dec 2025Complete.

Most important updates

  • 2 Dec 2025
  • 10 Jun 2025
  • 24 Jul 2024
  • 11 Dec 2023
  • 27 Sep 2023

Current obstacles

  • Compute gap with US labsMistral has a fraction of US labs' money and chips, so it competes on efficiency and flexibility.

Physics limits

  • All experts must sit in memoryMixture-of-experts saves arithmetic, not memory: a 1-trillion-parameter model needs about 1 TB at 8-bit precision even if only 3% runs per word, so it needs many linked chips.
  • Shrinking numbers loses informationRunning on smaller hardware means storing weights in 4 to 8 bits instead of 16. Below about 4 bits accuracy drops quickly, so model size sets a floor on memory needed.
  • Running out of human-written textFrontier models already train on 15 to 40 trillion tokens. Estimates put the usable stock of public human text at a few hundred trillion, so data, not chips, starts to cap plain scaling.

How it works

4 parts
Racks of a GPU cluster (CSIRO, Australia): Mistral trains its models on NVIDIA GPU clusters of this kind
Racks of a GPU cluster (CSIRO, Australia): Mistral trains its models on NVIDIA GPU clusters of this kindPhoto: CSIRO · CC BY 3.0 (opens commons.wikimedia.org)
Experts

Mixture of experts

Mixtral popularised open sparse models: 8 experts per layer, 2 used per word, so 47B parameters cost about as much to run as 13B.

Open

Open weights by default

Many models ship under Apache 2.0, so firms can run them on their own servers, which matters to European banks, governments and defence.

Sizes

From phones to data centres

The family spans small Ministral models for devices up to Mistral Large 3, a 675B-parameter mixture of experts with 41B active.

Reason

Magistral reasoning

Magistral models are trained with reinforcement learning to reason step by step, and can do it in European languages, not only English.

Update log

6 updates

Tue 2 Dec 2025

  • Minor: BlogSoftware

Tue 10 Jun 2025

  • Minor: BlogSoftware

Wed 24 Jul 2024

  • Minor: BlogSoftware

Mon 26 Feb 2024

  • Minor: BlogSoftware

Mon 11 Dec 2023

  • Major: BlogSoftware

Wed 27 Sep 2023

  • Major: BlogSoftware

About Mistral AI

The team behind Mistral models

Mistral AI

Open-weight models, Le Chat, AI cloud

Mistral AI is a Paris lab founded in 2023 by former Google DeepMind and Meta researchers. It makes efficient open and paid models and the Le Chat assistant. It is Europe's leading independent AI lab.

  • Raised about €3B led by Samsung in September 2026, valued above €21B (about $24B).
  • ASML led its €1.7B round in September 2025 and holds a significant minority stake.
  • The Mistral 3 family (December 2025) is open-weight under a permissive licence.

mistral.ai (opens mistral.ai)Company profile →

Founded
20233 yrs
Headquarters
France
Status
Private
Valuation
$24BprivateSep 2026
Raised
$5.6B3 rounds
Last round
Venture round · $3.5BSep 2026$24B post
Works in
AIOpen-weight modelsComputing
Coverage
2 programs · 11 updateslatest 8 Sep 2026checked 25 Sep
Lead investors
SamsungASMLLightspeed Venture Partners
Partners
NVIDIAMicrosoft
Investors
ASML
People
Arthur MenschCEO and co-founderGuillaume LampleChief Scientist and co-founderTimothée LacroixCTO and co-founder
Primer