Mistral previews its trillion-parameter Large 4 model
Mistral Large 4 is available through a preview API, with open weights promised by month-end. Its benchmark and efficiency claims still require independent testing.
AI / NEWS BRIEF
Beam targets coding, reasoning and agentic work with 23 billion active parameters. Early access is limited, while weights and technical materials are promised later in October.
Reflection AI has introduced Beam, its first open-weight model. Beam is a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active parameters, designed for coding, reasoning and agentic workloads. The company is offering limited early access while final red-teaming and evaluations continue. The weights are not yet public.
Reflection says it will release Beam’s weights under the Apache 2.0 licence later in October, together with a technical report, model card, documentation and developer tooling. The company reports pretraining on 23.8 trillion tokens and says a reinforcement-learning run generated more than 100 million rollouts using 10,500 Nvidia GB300 GPUs over four weeks.
Beam adds a substantial American entrant to a field of advanced open-weight models led largely by Asian developers. Reflection reports competitive coding and agent results with lower approximate inference compute than several larger systems, but those comparisons use estimated compute and mixed external data rather than measured end-to-end serving cost. The meaningful test will come after the promised artifacts are released and independent evaluators can reproduce performance, inspect licence terms and assess safety.
Mistral Large 4 is available through a preview API, with open weights promised by month-end. Its benchmark and efficiency claims still require independent testing.
OpenAI will test visual ads during image generation and add third-party conversion and brand-suitability tooling, deepening ChatGPT’s advertising business.