Reflection AI unveils Beam: 501B-parameter open MoE with 1M context

Reflection AI announced Beam, an open-weight 501B-parameter MoE (Apache 2.0, weights later this month) with a 1M-token context window. Only 23B parameters are active per token, but serving needs ~500GB of memory at 8-bit. The company claims GLM-5.2-level reasoning at 3–4x less inference compute.

• 501B total parameters, 23B active per token via MoE routing
• 256K context by default, 1M via midtraining extension
• SWE-bench Verified: 80.9 vs 70.7 for Nemotron 3 Ultra
• Pretrained on 23.8T tokens using 6,144 GB300 GPUs in under 4 weeks