What Mistral is best at
- EU businesses that need European AI on European terms
- Fast, cheap API workloads that do not need a giant model
- Open-weight releases you can self-host, down to 3B for on-device
Where it isn’t the answer
Bleeding-edge peak performance - it competes on efficiency and openness, not the crown.
Plans & pricing
The story so far
- Sep 2023Mistral 7B - tiny model, huge splash
- 2024Mixtral popularizes mixture-of-experts in the open
- Dec 2025Mistral Large 3: 675B parameters, 41B active, released as open weights
- Mar 16, 2026Small 4 folds reasoning, vision and coding into one model; Ministral 3 ships at 14B, 8B and 3B under Apache 2.0
- Apr 29, 2026Medium 3.5 becomes the flagship: frontier-class, multimodal, tuned for agents and coding
- 2026OCR 4.1, Voxtral speech models and Codestral extend the platform well beyond chat
Questions people actually ask
Why choose Mistral?
Efficiency and jurisdiction: strong models at low prices, from the only European frontier lab - a real answer to EU data-residency requirements.
What is the current Mistral flagship?
Mistral Medium 3.5, from April 2026 - Mistral’s own docs list it as the frontier-class model, tuned for agents and coding. Large 3 is bigger at 675B parameters but now sits with the open-weight models you run yourself.
Is Le Chat free?
Yes, with a generous allowance; Pro at around $15/month lifts the limits.
Are Mistral models open?
Several are released as open weights you can self-host, including the whole Ministral 3 family under Apache 2.0. The largest are API-only.