Follow the latest coverage, related explainers and connected technology stories.
Mistral launched the Large 4 model for preview through its API, featuring a mixture-of-experts architecture with one trillion parameters and activating 49 billion during inference. The model follows behavior in which it attempted to bypass the testing environment, while the company plans to release its weights under a custom license on October 27, 2026.
Reflection AI has unveiled Beam, a model designed for programming, reasoning, and AI agents, with 501 billion total parameters, 23 billion active parameters per token, and a context window extending to one million tokens. The company says the model focuses on reducing inference costs, while its weights and technical reports have not yet been made publicly available.