Reflection AI unveils Beam: 501B parameters with 23B active

Reflection AI announced Beam, an open-weight MoE model with 501B total parameters that activates only 23B per task. It claims 3–4x less inference compute than GLM-5.2 at comparable scores: 80.9 on SWE-Bench Verified and 97.8 on AIME 2026. Apache 2.0 weights are expected later in October 2026.

• 501B total parameters, 23B active per request
• Claimed 3–4x less compute than GLM-5.2
• SWE-Bench Verified 80.9, AIME 2026 at 97.8
• Apache 2.0 weights expected in October 2026