Mistral AI
Top European open-source models for sovereign AI
Mistral AI is the French startup founded in 2023 in Paris by former Google DeepMind and Meta researchers, which has become in 2026 the only European player capable of delivering frontier-class open-weights AI models. With a €11.7 billion valuation since its September 2025 Series C led by ASML and annual recurring revenue jumping from $20M to $400M between January 2025 and January 2026, the company has established itself as the credible European alternative to OpenAI, Anthropic and Google. The 2026 catalog now covers the full spectrum of use cases: Mistral Large 3 as the generalist flagship model, Mistral Small 4 (released March 16, 2026) which unifies reasoning, vision and coding in a single 119B-parameter MoE model with 256K token context, Magistral 1.2 for reasoning, Devstral 2 for agentic coding, the Ministral 3 family for edge and mobile deployments, Voxtral TTS for voice synthesis, and Codestral for IDE code completion. Most of these models ship under Apache 2.0 or modified MIT licenses, downloadable on HuggingFace and deployable on owned infrastructure. The offering breaks down into three channels: Le Chat (consumer interface, free plan with rare access to Mistral Large, Pro at €14.99/month, Team at €24.99/user/month), Mistral AI Studio (formerly La Plateforme, hosted API 5 to 10 times cheaper than GPT-5 or Claude Opus), and Mistral Forge (announced March 2026 at NVIDIA GTC, a platform for training custom models on proprietary data). Mistral is also building Mistral Compute, a Paris-area data center equipped with 13,800 NVIDIA Grace Blackwell GPUs powered by nuclear energy, to reduce its dependence on American infrastructure. Mistral now targets three distinct audiences: technical teams that want real control via self-hosting, organizations subject to sovereignty or regulatory compliance constraints (healthcare, defense, finance, public sector), and high-volume API developers for whom price-quality ratio matters more than absolute reasoning benchmarks. Its models still trail American flagships by one notch on the most demanding reasoning benchmarks (GPQA Diamond, AIME), a gap Mistral partly offsets with its specialized Magistral and Ministral 14B reasoning models.
View profile