Mistral AI releases Mistral Medium 3: 22B parameters, best-in-class cost-performance for European AI
In one sentence Mistral AI launches Medium 3, a 22-billion-parameter model delivering leading cost-to-performance for code, function calling, and JSON mode workloads, available on La Plateforme and as open weights.
Think of AI models like cars: you have the expensive high-end sports cars (massive, powerful, costly) and the cheap economy models (affordable but limited). Mistral Medium 3 is the premium mid-range sedan — near top-tier performance at a fraction of the price.
Mistral AI, the French startup that has become Europe's flagship AI company, has released its new 22-billion-parameter model. The pitch is straightforward: best cost-performance in its class. Medium 3 does not aim to beat GPT-4o or Claude Opus on every single benchmark, but it aims to deliver the best results per euro spent.
For developers and businesses, the improvements are practical. Medium 3 is significantly better at writing and understanding code, handles structured outputs like JSON reliably without extra prompt engineering, and supports function calling — the ability for AI to decide which tool or API to invoke inside an automated workflow — with greater accuracy than its predecessors.
For European companies, the data sovereignty angle is a major selling point. Medium 3 can be used through La Plateforme, Mistral's own cloud with European data centers, or downloaded and run entirely on your own servers. Either way, sensitive business data never needs to leave European jurisdiction, making GDPR compliance straightforward rather than an afterthought.
In short, a European company can now run a frontier-class AI model without routing data through American cloud providers, while keeping operational costs under control.
Companies
Mistral AI
Tools
Mistral Medium 3
Tags
Sources