Nvidia-backed Reflection AI unveils Beam open-weight model

Reflection AI debuted Beam, a 501-billion-parameter open-weight mixture-of-experts model with 23 billion active parameters, claiming performance competitive with Chinese models such as Z.ai GLM-5.2 on reasoning, coding, and agentic tasks while using substantially less inference compute. Weights are planned for release later in the month under Apache 2.0 after final red-teaming. The launch represents a Western open-weight response aimed at matching efficiency of leading Chinese open systems.