Reflection Previews Beam AI Coding Model, Optimized for Efficiency

Reflection AI unveiled Beam, its first model for coding and tool-using tasks. Beam has 501 billion total parameters (23 billion active), was pretrained on 23.8 trillion tokens and trained on 10,500 Nvidia GB300 GPUs. Downloadable weights under Apache 2.0 are promised for later in October.

• Beam scores 80.1 on Terminal Bench v2.1 vs 63.8 for Inkling
• On SWE Bench Pro v1 Beam gets 65.5 vs 54.3 for Inkling and 62.1 for GLM-5.2
• 501B total parameters, 23B active, pretrained on 23.8T tokens
• Apache 2.0 weights and documentation promised for late October