Reflection AI releases Beam, an open-weights model with 501 billion parameters; it lags behind leading open-source models on coding benchmarks but claims lower inference compute.

Reflection AI releases Beam, an open-weight model with 501 billion parameters, using a MoE architecture with 23 billion active parameters per token, and plans to publicly release the weights under the Apache 2.0 license later this month. On Terminal Bench v2.1, Beam scores 80.1, below GLM 5.2's 81.0, Kimi K3's 88.3, and DeepSeek V4.1 Flash's 90.6; on DeepSWE v1.1 it scores 44.4. Reflection states that Beam's advanced reasoning benchmarks are comparable to GLM 5.2, with three to four times less inference compute usage; this estimate is based on a formula and does not account for prompt processing and attention overhead. Safety evaluations and red teaming results will be released alongside the weights.