Reflection unveils Beam, an open-weight model aimed at Chinese rivals

Reflection debuted Beam on Monday, an open-weight text-only mixture-of-experts model with 501 billion total parameters and 23 billion active, plus a one-million-token context window. The company pitched it on inference efficiency against Chinese open models and promised Apache 2.0-licensed weights later in October.

The model itself isn't released yet — the weights are pending — and on the benchmark table Reflection published, Beam trails Kimi K3, GLM-5.3, and DeepSeek.

Sources: TechCrunch