Reflection AI, a Brooklyn-based startup backed by Nvidia, has introduced Beam as its first frontier open-weight model.

The company says Beam competes with mainstream open-source models on advanced reasoning benchmarks while dramatically reducing token costs and inference compute compared with rivals.

Beam is a text-only mixture-of-experts model trained with high‑compute reinforcement learning and is designed for reasoning, coding, and agent tasks.

Beam comprises 501 billion total parameters, with 23 billion active parameters. It was pretrained on 23.8 trillion tokens and features a context window of up to 1 million tokens.

Reflection notes that its performance claims have not been independently verified, but says Beam scores on par with Z-ai’s GLM-5.2 on advanced reasoning benchmarks, outperforms current mainstream Western open-source models, and requires three to four times less inference compute. The company positions Beam as a flagship model for enterprises, the public sector, and developers.