Reflection Launches Beam: 501B Open-Weight MoE Model
Original: Beam: Reflection's 501B open-weight model
Why This Matters
A 501B open-weight model with frontier inference efficiency raises the open-source compute bar.
Reflection AI unveiled Beam, a 501B-parameter sparse Mixture-of-Experts model with 23B active parameters, trained on 23.8T tokens with 100M+ RL rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks.
Beam is Reflection AI's first open-weight model, targeting coding, reasoning, and agentic workloads. The sparse MoE architecture activates 23B of its 501B total parameters at inference time, delivering efficiency comparable to much smaller models while matching larger ones on benchmarks.
Pretraining covered 23.8 trillion tokens from curated web and proprietary licensed datasets. The reinforcement learning phase generated over 100 million rollouts using 10,500 NVIDIA GB300 GPUs across four weeks — a significant compute commitment for an open-weight release.
On SWE Bench Pro v1, Beam scores 77.2, ahead of Nemotron 3 Ultra (56.9) but below GLM 5.3 (84.3) and Kimi K3 (88.2). On AIME 2026, it reaches 97.8. Reflection says Beam matches GLM-5.2 on advanced reasoning while using 3–4× less inference compute — the efficiency angle is the clearest differentiator from frontier models like Kimi K3, which still leads on raw scores.
Weights, a technical report, model card, and developer artifacts are scheduled for release later in October 2026. The model is currently undergoing final red-teaming, with early access sign-ups open now.