Mistral AI

Mistral AI is a Paris-based foundation-model lab that ships one of the broadest catalogues outside of OpenAI and Anthropic. The /models page lists generalist LLMs (Mistral Large 3, Medium 3.5, Small 4, Ministral 3-8B for edge), plus specialist models including Codestral and Devstral 2 for coding, Voxtral for TTS and transcription, and a Mistral OCR model the company pitches as the best document-extraction system on the market. Most flagship models are multimodal and tuned for agentic workflows.

What makes Mistral distinctive is the open-weight strategy: Small 4 ships under Apache 2.0 and Large 3 is offered as open-weight, while Medium and the specialists sit behind a premier API. That makes Mistral the default pick for teams that want a credible non-US frontier vendor, EU data residency, or the option to self-host the same model family they call via API. Pricing is per-token through la Plateforme, with enterprise contracts and on-prem deployments available for regulated industries.

The ecosystem now spans la Plateforme (API), Le Chat (consumer chat), fine-tuning, agents, and partnerships with Azure, AWS Bedrock, and Vertex. Caveats: the open-weight tier still trails GPT-5 and Claude on the hardest reasoning benchmarks, and the model lineup churns quickly enough that integration code needs revisiting every few months.