Mistral-Large-3

Mistral Large 3 is a state-of-the-art General-purpose Multimodal granular Mixture-of-Experts model with 39B active parameters, 673B total parameters featuring 128 experts per layer and Multi-Latent attention. Mistral Large 3 is an open-weight model optimized for long-context, multimodal, and instruction reliability. Mistral Large 3 stands in the leading tier of open models alongside DeepSeek, Kimi, Qwen 3, and GPT OSS. It shows clear strengths in instruction reliability, long-context comprehension, multimodal reasoning, and overall stability. While it is not designed to chase peak scores on abstract reasoning or math-heavy tasks, it delivers consistent quality across dialogue, knowledge, and applied reasoning workloads. Across a wide range of evaluations, Mistral Large 3 performs among the best open models for following instructions, sustaining multi-turn context, and maintaining coherence in long or complex exchanges. It handles extended inputs and multimodal content with steady accuracy, showing fewer breakdowns and more predictable results than most peers. The model’s balanced behavior makes it well-suited for production-grade assistants, retrieval-augmented systems, and multimodal applications. Within the global open-source landscape, Mistral Large 3 stands out as the strongest fully open model developed outside China. It offers frontier-level capability with Apache 2.0 licensing, reproducible results, and competitive performance against leading Chinese open models such as DeepSeek and Kimi. For organizations seeking a high-performance, open, and globally accessible alternative, Mistral Large 3 represents the benchmark for dependable frontier-class intelligence.