Reflection AI open-sources Beam, a 501B-parameter MoE for coding

Reflection AI has unveiled and open-sourced Beam, a 501B-parameter Mixture-of-Experts (MoE) model with 23B active parameters. Pretrained on 23.8T tokens and RL-tuned on 10.5K Nvidia GB300 GPUs, Beam demonstrates strong performance on coding benchmarks, matching Z.ai's GLM-5.2 while using significantly less inference compute. Weights and a model card are expected to ship later in October under Apache 2.0.