We advance the open frontier. Mistral is the pioneer of open artificial intelligence research, making seminal contributions across reasoning, multimodality, code, voice, and physical systems. Explore our model collection General purpose Models built for reasoning, language, multimodal understanding, and coding. Models built to excel at specific tasks and domains, including document intelligence, audio, robotics, and physical AI. Mistral Large 4 1T - 49B active Open-weight hybrid instruct-and-reasoning MoE with multimodal input; unifies instruction, reasoning, and agentic capabilities in a single model, state-of-the-art among open weights on cybersecurity, finance, and manufacturing, natively fluent in 160+ languages. Mistral Medium 3.5 Merged dense model unifying instruction, reasoning, and coding; configurable reasoning effort with multimodal input. Mistral Small 4 119B total · 6.5B active Hybrid instruct and reasoning MoE; unifies Instruct, Magistral, and Devstral into one model with multimodal input. Cascade Distillation for compact models; base, instruct and reasoning variants with image understanding. The world’s best document extraction and understanding model. Speech generation with reference voices. Voxtral Mini Transcribe Realtime Streaming transcription with configurable latency. Audio and text understanding; transcribes, translates, and answers questions from speech with text output. Mistral's time-series foundation models. State-of-the-art probabilistic forecasts, 5M–209M params, runs on CPU. 119B total · 6B active Proof engineering and verification in Lean 4. Text and image safety classification against a supplied policy. Read documentation Discover how our models and methods advance reasoning, perception and action. Filter by topic Foundation models Foundation models Policy-adaptive multimodal safety classification. Robostral Navigate Vision-only navigation across robot platforms. Phoenix-VL 1.5 Medium Localized multimodal models without capability loss. Expressive multilingual speech generation and adaptation. Voxtral Realtime Sub-second streaming speech recognition. Cascade-distilled models for constrained compute. Foundation models Going with the Speed of Sound Neural surrogates for transonic aerodynamics. Fluid Intelligence Compute scaling laws for CFD. AB-UPT for Automotive and Aerospace

Fast neural surrogate aerodynamics. 5D neural surrogates for plasma turbulence. Small models designed for coding agents. Audio understanding with strong text capabilities. Pure reinforcement learning for reasoning. Scalable neural automotive CFD surrogates. Real-time neural industrial particulate flow simulation. Natural-resolution image and document understanding. Universal Physics Transformers Unified neural operators for physics. Mixtral of Experts Frontier performance through sparse MoE. Foundation models Reset expectations for small, efficient language models. Foundation models Try adjusting your search or filters to find what you are looking for. Research areas. Reasoning and multimodality Reason across language, media, and code. Document intelligence. Models that excel at parsing complex documents. Multilingual speech understanding, generation, and adaptation. Safety and guardrailing. Models that evaluate text and images against policies written in natural language. Complex environment simulation, from automotive airflow and plasma turbulence to industrial particle flows. Navigation models that follow language instructions and work across robot types using a single RGB camera. Inside the work. Go deeper on the techniques, systems, and engineering behind our models—from training and evaluation to inference and efficiency. Introducing Mistral Large 4 Introducing Mistral OCR 4 State of the art document intelligence model. Introducing Robostral Navigate Robostral Navigate, our first model built for embodied navigation. Leanstral 1.5: Proof Abundance for All Introducing Shieldstral. Physics AI research that’s shaping the industry. Published breakthroughs pushing the state of the art. Speaking of Voxtral Voxtral TTS: A frontier, open-weights text-to-speech model that’s fast, instantly adaptable, and produces lifelike speech for voice agents. Build on our open-weight models. Download, adapt, and deploy our models. Discover Mistral on Hugging Face Discover Mistral on Github We’re seeking researchers and engineers to advance the models, systems, and techniques behind Mistral. Explore open roles