[AINews] Reflection Beam - 501B-A23B American Open Model
Reflection AI launches Beam, a 501B/23B-active US-trained open MoE model for coding and agentic work
Reflection AI has ended its stealth period by launching Beam, a 501B-total/23B-active Mixture-of-Experts model trained from scratch in the US, targeting coding, agentic, and scientific workloads. While benchmarks place it behind frontier open models like GLM 5.3, Kimi K3, and DeepSeek V4.1 Flash, it addresses a market demand for domestically-trained open-weight alternatives. The launch establishes Reflection as a functional neolab and adds a new US-origin option to the competitive open model landscape.