Reflection AI debuts Beam, an open-weight coding and agent model

Reflection AI presents Beam as a major open-weight frontier release: an Apache-2.0 501B mixture-of-experts model with 23B active parameters, roughly 1M-token context, and coding, reasoning, and agentic positioning. The company reports 80.9 on SWE-bench, parity with GLM-5.2, and 3–4x lower inference compute, but the newly repeated coverage adds no independent validation and complete public artifacts remain pending.