Reflection AI Releases Beam, a 501B Open-Weight MoE Model

Reflection AI has released Beam, a sparse mixture-of-experts model with 501 billion total parameters but only 23 billion active per token, released as open weights under Apache 2.0. Beam was pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 GPUs in under four weeks, with a one-million-token context window. A separate reinforcement learning run used 10,500 GB300 GPUs for four weeks, generating over 100 million rollouts across nearly one million environments and about 1.3 billion sandboxes, with no sign of a plateau at the end of that run. Reflection claims a three-to-four times inference compute efficiency advantage over GLM-5.2, though one source notes this is based on estimated forward-pass FLOPs rather than measured end-to-end serving cost. On Terminal Bench v2.1, Beam scores 80.1, close to GLM-5.2's 81.0, but trails DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3, two models that do not appear in Reflection's headline benchmark chart. FP8 and NVFP4 quantized builds are also planned.