Reflection releases Beam, an open-weight reasoning model

Brooklyn-based startup Reflection AI has officially unveiled Beam, its first frontier, open-weight artificial intelligence model. The text-only system uses a mixture-of-experts architecture — a design where the model routes each token to specialized sub-networks — and is trained on high-compute reinforcement learning.

Model architecture and specifications

Beam features 501 billion total parameters with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and includes a 1 million token context window. According to the company, the model is designed to handle reasoning, coding, and agentic tasks while using significantly less inference compute than comparable systems.

Performance and positioning

On advanced reasoning benchmarks, Reflection reports that Beam scores on par with Z.ai's GLM-5.2 model while outperforming several other open alternatives. The startup intends for Beam to serve as a workhorse model for enterprises, developers, and the public sector.

Infrastructure and future releases