Summary
On October 5, 2026, Reflection AI announced Beam, its first open-weight model: a 501-billion-parameter mixture-of-experts model (23B active per token, 1M-token context) aimed at coding and agentic tasks, licensed Apache 2.0, with weights and a technical report promised later in October and early access via a waitlist.
What changed
Reflection AI unveiled Beam, a 501B-parameter MoE model activating 23B parameters per token with a 1M-token context window, positioned for coding and agentic use and released under Apache 2.0. At announcement the weights were not yet downloadable; Reflection said it would publish the weights, technical report, model card, and developer artifacts later in the month, with initial access gated behind a waitlist.
Why it matters
Beam is pitched as a Western open-weight frontier model competing on inference efficiency rather than peak capability, with Reflection claiming parity with GLM 5.2 reasoning at 3-4x less inference compute. For teams building agents and coding tools, an Apache-2.0 model that trades a little capability for much cheaper inference is attractive, though the delayed weight release and self-reported benchmarks temper how much can be validated today.
Evidence excerpt
Beam is a 501B-parameter mixture-of-experts model activating 23B parameters per token with a 1M-token context, pretrained on 23.8T tokens; Reflection says it matches GLM 5.2 on reasoning while using 3-4x less inference compute. Weights, technical report, and model card are to follow later in October under Apache 2.0; access is waitlisted at launch.