技术规格

系统要求

不同量化方法和上下文大小的显存要求

Mistral Large 3

Mistral Large 3 is a state-of-the-art, general-purpose multimodal model, characterized by its granular Mixture-of-Experts (MoE) architecture. The model incorporates 41 billion active parameters within a total parameter pool of 675 billion, representing a substantial computational capacity. It was trained from scratch using a cluster of 3000 NVIDIA H200 GPUs.

This variant is specifically an instruct-post-trained version, fine-tuned for instruction-following tasks. It is designed to excel in conversational AI, agentic functions, and other instruction-based use cases, making it suitable for deployment in production-grade assistants, retrieval-augmented generation (RAG) systems, scientific applications, and complex enterprise workflows. The model is engineered for reliability and robust long-context comprehension, supporting a context window of 256,000 tokens.

A key architectural component of Mistral Large 3 is its integrated 2.5 billion parameter Vision Encoder, enabling multimodal capabilities that allow the model to analyze images and derive insights from visual content alongside text. It offers strong multilingual support across dozens of languages, including major European and Asian languages. Furthermore, Mistral Large 3 demonstrates strong adherence to system prompts and provides best-in-class agentic capabilities, including native function calling and JSON output generation.

没有可用的 Mistral Large 3 评估基准。