Mistral

Mistral AI is a French AI company founded in April 2023 by Arthur Mensch, Guillaume Lample, and Timothée Lacroix — all former researchers from Google DeepMind and Meta. Headquartered in Paris, Mistral quickly became Europe's most prominent LLM lab, reaching a $14 billion valuation by 2025 after raising over €2.8 billion across seed, Series A, B, and C rounds.

Mistral's identity is built on a commitment to open-weight models alongside a commercial API, giving developers a genuine choice between self-hosting and managed access.

Open-weight model family (Apache 2.0 unless noted)

• Mistral 7B — the original 7B dense model that launched the company in September 2023; 32K context window; runs on consumer hardware; arguably the most influential small open LLM release of 2023.
• Mixtral 8x7B — Mixture-of-Experts architecture with 47B total parameters but only 13B active per token; 32K context window; matches or exceeds GPT-3.5 on most benchmarks at far lower inference cost.
• Mixtral 8x22B — larger MoE with 141B total / 39B active parameters; 64K context window; strong on coding and mathematics; faster inference than dense 70B models.
• Mistral NeMo — 12B dense model built in collaboration with NVIDIA; 128K context window; multilingual strength; Apache 2.0.
• Codestral — code-specialist model with fill-in-the-middle (FIM) support; designed for IDE integrations and agentic coding workflows; available via API ($0.30/$0.90 per 1M tokens) and as an open-weight download.
• Mathstral — 7B model fine-tuned for mathematical reasoning; Apache 2.0; optimised for chain-of-thought math problem solving.
• Mistral Small (open) — efficient small model available for self-hosting; suitable for classification, summarisation, and function calling without expensive GPU requirements.
• Mistral Small 4 (2026) — 119B total / ~6.5B active parameters per token; 256K context window; unifies reasoning, multimodal, and agentic coding in a single model with configurable reasoning effort.

Proprietary API-only models (via la Plateforme)

• Mistral Large / Mistral Large 2 — flagship dense model; 128K context window; strong on instruction following, function calling, and multilingual tasks; priced at $2.00/$6.00 per 1M tokens.
• Mistral Large 3 — frontier MoE with 675B total / 41B active parameters; 256K context window; available on la Plateforme and major cloud marketplaces.
• Mistral Medium / Mistral Medium 3.5 — mid-tier dense model (128B parameters); 256K context window; balanced cost/performance; powers Mistral's Vibe remote coding agents.
• Pixtral Large — multimodal (vision + text) flagship model; available API-only; $2.00/$6.00 per 1M tokens.
• Ministral 3B / Ministral 8B — ultra-efficient models targeting edge and on-device deployment.

Access methods

Self-hosting is the defining advantage of Mistral's open-weight models. Weights are published on Hugging Face under Apache 2.0 and can be run via:

• Ollama — single command (ollama pull mistral) for local inference on Mac, Linux, or Windows
• vLLM — production-grade serving with continuous batching and OpenAI-compatible API
• llama.cpp — quantised inference on CPU and Apple Silicon with minimal VRAM requirements
• mistral-inference — Mistral's official Python inference library (10.5K+ GitHub stars)

Mistral API (la Plateforme) at api.mistral.ai provides OpenAI-compatible REST endpoints for both open and proprietary models. The API offers a free tier for experimentation, function calling, JSON mode, and streaming. Pricing starts at $0.04/1M tokens for Ministral 3B and scales to $6.00/1M output tokens for Mistral Large and Pixtral Large.

Third-party cloud deployments are available on Amazon Bedrock, Microsoft Azure AI Foundry, Google Cloud Vertex AI, and Together AI — enabling enterprises to access Mistral models within existing cloud contracts and compliance boundaries, including EU-region hosting (Paris data centre, AWS Frankfurt, Azure EU regions).