Reading up on Beam

Reflection unveils Beam, a 501B‑parameter sparse MoE model with 23B active parameters, trained on 23.8T tokens and scaled RL, delivering frontier coding and reasoning performance at high inference efficiency.