Reflection AI Releases Beam-1.0, a 501-Billion-Parameter Open-Weight Model Competing on Agent Tasks

Reflection AI, a Brooklyn-based AI startup founded by former Google DeepMind researchers, released Beam-1.0-501B-A23B on October 16 under the Apache 2.0 license, marking a significant milestone for open-weight models in reasoning, coding, and agentic workloads. The model combines 501 billion total parameters with a sparse mixture-of-experts architecture that keeps only 23.4 billion parameters active per token, reducing inference compute by roughly 3–4× compared to equivalent dense models.