Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, but only 23 billion active during each step. It is built for coding, reasoning, and agentic workloads, so it can handle long, complex tasks while keeping compute and latency practical. The sparse architecture gives you the broad knowledge of a much larger model without paying the full inference cost on every single request.
Launch Team / Built with
MI
Product Info
Socials
Launched October 24, 2026
Awards