Reflection AI releases Beam, a 501B open-weight MoE model

Reflection AI released Beam on October 5, 2026: a sparse MoE model with 501.3B total and 23.4B active parameters under Apache 2.0. It uses 3–4× fewer inference tokens than comparable open models while matching GLM-5.2 on solution quality.

• 501.3B total parameters, 23.4B active per token
• Pretrained on 23.8T tokens, RL over 100M rollouts on 10,500 GB300 GPUs
• SWE-bench Verified 81.6%, GPQA Diamond 90.6%
• DFlash2 draft model speeds generation 1.72–3.03×