Mistral Large 3
Overview
Mistral Large 3 is a general-purpose, open-weight, multimodal language model developed by Mistral AI. It employs a granular Mixture-of-Experts (MoE) architecture, featuring 41 billion active parameters and a total of 675 billion parameters. Trained from scratch on 3,000 NVIDIA H200 GPUs, it is designed to handle long-context comprehension, with a context window of up to 256,000 tokens. The model is released under the Apache 2.0 license, facilitating open access and customization for developers and enterprises. Technically, Mistral Large 3's MoE architecture allows for dynamic selection of active experts during inference, optimizing computational efficiency and enabling scalability. This design contributes to its strong performance across various tasks, including instruction following, multimodal reasoning, and coding assistance. The model's multilingual capabilities support over 40 languages, and it demonstrates proficiency in both text and image understanding. Mistral Large 3 is positioned as a competitive open-source alternative to proprietary models, offering high performance with the flexibility of open weights. Mistral Large 3 is intended for enterprise-grade applications, including long document understanding, AI assistants, agentic and tool-use capabilities, enterprise knowledge work, and general coding assistance. It is engineered for production-grade assistants, retrieval-augmented systems, scientific workloads, and complex enterprise workflows, providing powerful long-context performance and stable cross-domain behavior. The model is available through various platforms, including Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, Modal, IBM WatsonX, OpenRouter, Fireworks, Unsloth AI, and Together AI, with plans for future availability on NVIDIA NIM and AWS SageMaker. Mistral Large 3's performance is competitive within the open-source landscape, achieving notable results on several benchmarks. For instance, it scored 85.5 on the MMLU Massive Multitask Language Understanding benchmark, 92.0 on the HumanEval coding benchmark, and 93.6 on the MATH-500 math benchmark. These scores reflect its capabilities in reasoning, coding, and mathematical problem-solving. The model's strengths include coding, long context, reasoning, and math, making it suitable for a wide range of applications requiring advanced language understanding and generation capabilities.