Mistral

startup

france · Founded 2023 · Valuation: $14B as of 2025-09

Mistral AI is a French AI company founded in April 2023 by three Ecole Polytechnique alumni: Arthur Mensch (CEO, ex-Google DeepMind), Guillaume Lample (Chief Scientist, ex-Meta FAIR), and Timothee Lacroix (CTO, ex-Meta FAIR, contributed to LLaMA). With $3B+ raised across five equity rounds (seed through Series C) and valued at ~$14B, Mistral is Europe's leading AI lab. Revenue surged from ~$20M ARR in January 2025 to ~$400M ARR by January 2026.

Mistral pioneered efficient open-weight models: Mistral 7B (Sep 2023) introduced Sliding Window Attention and beat Llama 2 13B; Mixtral 8x7B (Dec 2023) popularized Sparse MoE for open models; Mixtral 8x22B scaled to 141B/39B active. The current flagship Mistral Large 3 (Dec 2025, 675B/41B active MoE, Apache 2.0) competes with frontier models across reasoning, coding, and multilingual tasks. Mistral Small 4 (Mar 2026, 119B/6.5B active MoE) unifies instruct, reasoning, multimodal, and agentic coding in one model.

Beyond LLMs, Mistral has built a full product ecosystem: Le Chat (consumer AI assistant), Codestral (code-specialized models including the Mamba2-based Codestral Mamba), Pixtral (vision-language models), Voxtral (speech understanding and TTS), Magistral (pure RL reasoning), Mistral Saba (Arabic/South Asian languages), and Devstral (agentic coding, 72.2% SWE-Bench Verified). Mistral is building a 13,800 GB300 GPU data center near Paris ($830M debt financing, Mar 2026) and holds defense and enterprise contracts including with the French military and CMA CGM.

Website Wikipedia HuggingFace GitHub Artificial Analysis OpenRouter
frontieropen-weightmoemultimodalreasoningcoding

People

News

Outputs (16)

Mistral Small 4

model

<p>119B total / 6.5B active MoE (128 experts, top-4 routing). First Mistral model to unify instruct, reasoning, multimodal, and agentic coding in one…

Mistral Large 3

model

<p>Mistral's largest model. 675B total / 41B active MoE with granular mixture-of-experts architecture, plus 2.5B vision encoder for native multimodal.…

Ministral 3

model

Edge-optimized model family (3B, 8B, 14B) created via Cascade Distillation from Mistral Medium 3.1. 3 variants each (base, instruct, reasoning) = 9 mo…

Voxtral

model

Speech model family. Voxtral Mini (3B) and Small (24B) handle long-form audio understanding (30-40 minutes). Voxtral TTS (4B) adds zero-shot voice clo…

Magistral

model

<p>Reasoning model family trained with <strong>pure reinforcement learning</strong> — no distillation from external reasoning models. ~50% boost in AI…

Devstral 2

model

Agentic coding model family. Devstral 2 (123B, Dec 2025) achieves 72.2% SWE-Bench Verified. Devstral Small 2 (24B) achieves 68.0%. Runs on single RTX…

Mistral Saba

model

24B regional language model specialized for Arabic and South Asian languages (including Tamil and other South Indian languages). More accurate than mo…

Pixtral 12B

model

12B multimodal model with 400M vision encoder trained from scratch. Processes images at natural resolution and aspect ratio. MMMU: 52.5%, DocVQA: 90.7…

Mistral Large 2

model

123B dense model with 128K context. Performs on par with GPT-4o, Claude 3 Opus, and Llama 3 405B. MMLU: 84.0%. Supports 80+ programming languages and…

Mistral NeMo

model

12B dense model built in collaboration with NVIDIA. 128K context. MMLU: 68.0%. Outperforms Gemma 2 9B and Llama 3 8B. Apache 2.0.

Codestral Mamba

model

7.3B code model using Mamba2 state-space architecture (NOT a transformer). Linear-time inference with theoretically unlimited context (tested to 256K…

Mathstral 7B

model

7B math-specialized model based on Mistral 7B. MATH: 56.6% (single pass), 68.37% (majority voting), 74.59% (with reward model among 64 candidates). Ap…

Codestral

model

22B code-specialized model supporting 80+ programming languages with fill-in-the-middle (FIM). 32K context. Strong on RepoBench long-range code genera…

Mixtral 8x22B

model

Scaled MoE architecture to 141B total / 39B active per token (8 experts, top-2 routing). 64K context. MMLU: 77.3%. Outperforms Command R+ and Llama 2…

Mixtral 8x7B

model

<p>Popularized Sparse Mixture-of-Experts for open-weight models. 46.7B total / 12.9B active per token, with 8 experts per layer and top-2 routing. 32K…

Mistral 7B

model

<p>Mistral's debut model and a landmark in efficient open-weight LLMs. 7.3B dense parameters with two key innovations: <strong>Grouped-Query Attention…