Mistral Large 4

The cost of using advanced AI models drops by 50% for output tokens and 50% for input tokens.

Mistral AI has announced the public preview of Mistral Large 4, a multimodal model featuring a granular Mixture-of-Experts architecture. The model boasts 49 billion active parameters and a total of 1.05 trillion parameters, complemented by a 1.6 billion parameter vision encoder. Mistral Large 4 supports a 1 million token context window and is designed for general-purpose use. Pricing details indicate input tokens at $1.36 per million, cached input at $0.68 per million, and output tokens at $4.18 per million. The model is available through an API, with features including structured outputs, function calling, document QnA, prefix operations, chat completions, batching, agents, conversations, and built-in tools. This release positions Mistral AI to compete in the advanced AI model market.

Mistral Large 4's architecture, with its high total parameter count and active parameter distribution, signals a continued trend towards efficient scaling in large language models. The inclusion of a dedicated vision encoder and a substantial context window caters to the growing demand for multimodal capabilities, enabling more complex reasoning and interaction with diverse data types. This release challenges established players by offering competitive performance and advanced features, potentially driving further innovation in model design and application.

The competitive pricing structure, particularly the significant reduction in output token costs, could democratize access to powerful AI for a wider range of developers and businesses. This encourages experimentation and the development of new AI-powered products and services that were previously cost-prohibitive. The broad feature set, including agents and built-in tools, suggests a move towards more integrated and autonomous AI systems, requiring developers to adapt their integration strategies.