Mistral AI

Overview

Paris-based AI research and product company. Open-weight model lineup (Mistral, Mixtral) + Le Chat chatbot + Vibe coding tools. A leading frontier player in the European camp.

Key People

TBD

Models & Products (2026)

• Mistral Large 3 — ~2026-07-06 early access, ~675B MoE (41B active), Apache 2.0, "fat but sparse"; GA expected mid-to-late July → Mistral Large 3
• Robostral Navigate — 2026-07-08, 8B robotics navigation, single RGB camera, 76.6% R2R-CE (new SOTA, beats depth+multi-camera systems), sim-only training
• Leanstral 1.5 — 2026-07-01/02, 119B MoE (6.5B active), Apache 2.0, Lean 4 formal verification, 100% miniF2F, 587/672 PutnamBench
• Mistral OCR 4 — 2026-06-23, structure-aware document AI, self-hosted single container, 170 languages, OlmOCRBench 85.20 (top score), 72% annotator win rate, $4/$2 per 1K pages (API/Batch) (source)
• Devstral 2 — 2026-05, 123B (SWE-bench 72.2%) + 24B Small variant; powers Vibe CLI
• Mistral Vibe CLI — open-source CLI coding agent, Apache 2.0 (github.com/mistralai/mistral-vibe)
• Mistral Medium 3.5 — 2026-04-29
• Mistral Forge — 2026-03-26, enterprise model training platform on proprietary data
• Le Chat — chatbot (now with "Work mode" added)

Recent Activity

• 2026-07-21: Mistral + Microsoft Foundry "multibillion" expansion — Vera Rubin GPUs, Azure Local sovereign deployment — Microsoft and Mistral announced a major strategic partnership expansion on July 21, 2026. Key terms: Mistral Medium 3.5 and Mistral OCR 4 now available in Microsoft Foundry and Copilot Studio; Mistral models deployable via Azure Local (sovereign/air-gapped on-premise, no data egress); workloads running on Vera Rubin GPU infrastructure in European data centers. Deal scale: described as "multibillion" (exact terms not disclosed). Why it matters: Vera Rubin GPU availability in European DCs gives EU enterprise customers next-generation NVIDIA compute at locations that satisfy GDPR and EU AI Act data-residency requirements. Azure Local deployment extends Mistral's sovereign-AI pitch from open-weights to sovereign cloud: organizations can now run frontier Mistral models entirely within their own infrastructure boundary. This directly reinforces Mistral's EU digital sovereignty positioning vs. US-hosted OpenAI/Anthropic/Google models — a key differentiator in European public-sector procurement. The Foundry + Copilot Studio availability immediately puts Mistral into Microsoft's enterprise distribution channel without requiring direct Mistral sales motion. → (source) (Microsoft)

• 2026-07-06: Mistral Large 3 MoE enters early access — A ~675B parameter MoE model (41B active per token), Apache 2.0 licensed, opened early access to developers on July 6. CEO Arthur Mensch confirmed the model July 4 in a TechCrunch profile, describing the architecture as "fat but sparse" — high total capacity with sparse per-token activation, combining frontier-class capability with inference efficiency. Benchmarks not yet published (pending GA). Broader release expected mid-to-late July 2026. Why it matters: at ~675B/41B active with Apache 2.0, this would be the largest open-weight MoE available under a permissive license — directly challenging DeepSeek-V3 (671B/37B active, MIT) and Llama's open-weight leadership. EU enterprise users get a sovereignty-compliant frontier-class model with no usage restrictions. → Mistral Large 3 (source) (TechCrunch)

• 2026-07-08: Robostral Navigate — Mistral's first robotics model — An 8B robotics navigation model that lets robots navigate complex environments using only a single RGB camera and plain-language instructions. Hardware-agnostic; trained entirely in simulation (CISPO online RL). Achieves 76.6% on R2R-CE (Room-to-Room in Continuous Environments) validation unseen — new state of the art, beating prior best single-camera approach by +9.7 points and prior best with depth/multi-camera by +4.5 points. Applications: manufacturing, delivery, logistics, hospitality. Why it matters: Mistral's first physical AI product, following the Emmi AI (physics simulation) acquisition in May. Establishes a third axis beyond language models and coding tools — and immediately achieves SOTA despite sim-only training and single-camera constraints, lowering the hardware bar for robot deployment. → Robostral Navigate (source) (Bloomberg)

• 2026-07-04: CEO confirms open-weight "fat but sparse" MoE for summer, ARR tops $400M — In a TechCrunch profile, CEO Arthur Mensch confirmed a forthcoming summer model: "We have a very exciting model to come this summer — it will be open-weight, and we're opening early access to it in July." Mensch described the architecture as "fat but sparse" — a large total-parameter MoE (high capacity) with sparse activation per token (low per-token compute cost), combining frontier capability with inference efficiency. ARR has climbed from $20M (one year prior) to above $400M, on track to exceed $1B by end of 2026. Mensch also reiterated that closed AI models "give providers immense leverage" over enterprise customers — positioning Mistral's open-weight strategy as a sovereignty guarantee. The exact parameter count, benchmarks, and license are not yet disclosed. Why it matters: a "fat but sparse" MoE targeted at European open-weight users would directly compete with Llama's open-weights position and DeepSeek-V3's MoE efficiency; EU-based enterprises gain a credible sovereignty alternative. Early access expected in July 2026. (source) (TechCrunch)

• 2026-07-01/02: Leanstral 1.5 — open-source formal verification model saturates miniF2F at 100% — Mistral released Leanstral 1.5 (July 1–2, 2026), a 119B MoE model (6.5B active, Apache 2.0) purpose-built for Lean 4 formal verification and automated theorem proving. Benchmarks: 100% miniF2F (full saturation), 587/672 PutnamBench (87.4%), FATE-H 87%, FATE-X 34%. Agentic proof engineering: found 5 previously unknown bugs across 57 open-source repositories. Available as free API endpoint (leanstral-1-5) and HuggingFace weights. Why it matters: formal verification has historically required specialized expertise; a production-ready open-source model that saturates the leading benchmark makes mathematically certified software accessible to any developer with access to the weights. The Apache 2.0 license enables commercial deployment in safety-critical domains (aerospace, finance, infrastructure). → Leanstral 1.5 (source) (Mistral) (The Decoder)

• 2026-06-23: Mistral OCR 4 — structure-aware document AI (self-hosted, 170 languages) — Mistral released OCR 4, a document intelligence model producing structured citation-ready output for enterprise RAG, search, and retrieval pipelines. Outputs: bounding boxes, block classification, and inline confidence scores alongside extracted text. 170 languages across 10 language groups. Runs in a single container (fully self-hosted; no cloud required; enterprise data sovereignty). Benchmarks: OlmOCRBench 85.20 (top score across all systems tested); 72% independent annotator win rate vs. all leading OCR systems. Pricing: $4/$2 per 1,000 pages (API/Batch); $5/1K for Document AI tier. Use cases: enterprise search ingestion, RAG pipeline document preparation, domain-specific retrieval (legal, financial, medical), regulated-industry self-hosting. Competing with Azure Document Intelligence, AWS Textract, Google Document AI. Differentiation: structured output + self-hosted + competitive accuracy. (source) (Mistral)

• 2026-05-23 (official blog, supplementary): "Introducing Physics AI at Mistral" — official detailed blog on the Emmi acquisition. Physics AI definition: "learn from physics-solver outputs → directly predict physical fields from geometry + boundary conditions, processed in a few seconds on a single GPU." Newly named customers: Safran (aerospace propulsion and equipment), Siemens Energy (power generation and grids) — added to the AI Now Summit announcements (Airbus, ASML). Technical value: replaces CFD/FEA simulations that previously took days with real-time processing. Optimizes digital twin asset operations. (source)

• 2026-05-28: AI Now Summit 2026 — combined announcements (source)

  • Vibe integrated agent launch: a single agent handles email and calendar, deep research, document writing, and orchestrating repetitive processes + coding (auto-generating up to PRs). A unified upgrade of the former Vibe CLI + Vibe remote agents.
  • Search Toolkit launch: a toolkit for production search pipelines, supporting cloud, on-premises, and hybrid.
  • Les Ulis compute facility (new): a 10MW inference-dedicated data center in Essonne (France), scheduled to open in Q3 2026. Mistral's first self-owned inference infrastructure — direct control over supply-chain risk.
  • Airbus industrial AI partnership: integrating AI across all of aircraft, helicopters, defense, and space. The first major case of the physics AI + engineering AI strategy tied to the Emmi acquisition.

• 2026-05-22: Emmi AI acquisition — expansion into Physics AI + industrial AI — acquired Vienna, Austria-based Emmi AI (NXAI spin-out, founded 2024). Specialized in Physics AI — computational fluid dynamics (CFD), heat transfer, material stress simulation. Targeting aerospace, automotive, energy, and semiconductor industries. All 30+ researchers and engineers join the Mistral Science & Applied AI team. Adds an official office in Linz (Austria). Terms undisclosed. → Entry into physics-simulation AI beyond language models + coding, strengthening European industrial AI positioning. (source)

• 2026-05-18 (ingest): Workflows capture + NVIDIA Nemotron Coalition recorded

• 2026-04-28: Mistral Workflows public preview — a durable AI orchestration engine based on Temporal. Resumes multi-step LLM pipelines (API calls + tool use + human input) even after crashes and restarts. Hybrid deployment (Mistral hosts the orchestrator, workflow code runs in the customer environment). Early users: ASML, CMA-CGM, La Banque Postale, etc. → Agents (LLM Agents) (source)

• 2026-03-16: NVIDIA Nemotron Coalition founding member — co-developing open frontier models with NVIDIA. Other members: Black Forest Labs, Cursor, LangChain, Perplexity, Reflection AI. Mistral contributes technology to training Nemotron 4 family-based models.

• 2026-05-17 (extended ingest): Mistral Connectors/MCP Studio capture (source)

• 2026-05-17 (ingest): Devstral 2 + Vibe CLI officially documented

• 2026-05 (announced): Devstral 2 + Mistral Vibe CLI — Devstral 2 (123B, 72.2% SWE-bench), Devstral Small 2 (24B, 68% SWE-bench), Vibe CLI (Apache 2.0 open-source coding agent). Self-reported claim: 7x cost efficiency versus Claude Sonnet (source)

• 2026-03-26: Mistral Forge — a platform for training frontier-grade models on enterprise proprietary data. Supports MoE architecture, integrated with Vibe CLI. Partners: ASML, ESA, Ericsson, etc. (https://mistral.ai/news/forge)

• 2026-04-15: Mistral Connectors + MCP in Studio — GitHub, Gmail, web search built in by default. Code Interpreter, Image Gen, Document Library tools. Custom MCP API support (applied to all model and agent calls). Human-in-the-loop approval flow. Applied across Le Chat + AI Studio. → Completes a fully open-source agent stack together with the Vibe CLI launch (source)

• 2026-04-29: Mistral Medium 3.5 + Vibe remote agents + Le Chat Work mode (source)

Strategic Position

• Physical AI (2026-07): Robostral Navigate (July 8) — first robotics model, SOTA R2R-CE single-camera navigation — + Emmi AI acquisition (May 22, physics simulation) = Mistral building a third axis beyond language + coding, targeting European industrial AI
• Central to the European AI sovereignty narrative
• Open-weight + commercial product dual track
• Entry into the agents space (tracking Agentic Reinforcement Learning, Agents (LLM Agents))
• Enterprise full stack: Devstral 2 (model) + Vibe CLI (coding) + Connectors/MCP (integration) + Workflows (orchestration) = a fully open-source agent stack
• NVIDIA Nemotron Coalition: securing access to compute and data pipelines through co-development of open frontier models
• Physics AI + industrial AI (new 2026-05): the Emmi AI acquisition (5/22) secures CFD, heat transfer, and material simulation capabilities. The Airbus partnership (5/28) secures a major aerospace customer. A differentiated position targeting European industry (existing customers such as ASML, ESA, Ericsson). The clearest point of differentiation from US labs.
• Self-owned compute infrastructure (new 2026-05-28): the Les Ulis 10MW facility reduces dependence on cloud supply chains. Scheduled to open in Q3 2026. Aligned with the European AI sovereignty narrative.

Related

• Devstral 2
• Mistral Medium 3.5
• Grok Build — competing coding CLI (xAI)
• Agents (LLM Agents) — Devstral/Vibe are in the agentic coding space

Conflicting Reports

None