French AI startup Mistral has launched the third version of its open-weight model series, featuring a large multimodal and multilingual language model (LLM) alongside nine smaller models tailored for specific tasks. The flagship Large 3 model employs a Granular Mixture of Experts architecture and boasts a 256,000-token context window. Mistral, founded by alumni from DeepMind and Meta, aims to challenge U.S. and Chinese AI leaders by positioning itself as a major European player. The company has announced partnerships with Singapore's Home Team Science and Technology Agency, German defense startup Helsing, and Stellantis to integrate its compact models into robots, drones, and vehicles.