Reflection releases Beam: a 501B-A23B open-source American model.

Reflection releases Beam, a text MoE model with 501B total parameters and 23B active, targeting coding, agent, and research tasks, trained from scratch, with full weights to be released under Apache 2.0 this month. Its pretraining used 23.8T tokens, with about four weeks of pretraining and four weeks of RL on approximately 10,500 GB300 units, claiming a SWE-bench Verified score of 80.9. Artificial Analysis expects it to be one of the most token-efficient open-source models at the same intelligence level, but some benchmarks still lag behind models like DeepSeek V4 Flash.