Beam, the open-weight model for coding and agents
Beam is a sparse Mixture-of-Experts model from Reflection AI, built for coding, reasoning and agentic workloads. Weights are scheduled for release under Apache 2.0. This site collects specs, benchmarks and practical guides for running it.
Key specs
Total parameters
501B
Active parameters
23B
Context length
1M
License
Apache 2.0
Pretraining tokens
23.8T
Architecture
MoE
Benchmark highlights
SWE-Bench Verified
80.9
AIME 2026
97.8
AA-LCR
79.3
GPQA Diamond
90.5