Reflection AI Releases Beam, a 501B Parameter Open-Weight Model
Reflection AI, founded by former DeepMind researchers, has released Beam, its first open-weight AI model. Designed specifically for coding, reasoning, and agentic workloads, Beam is claimed to be 3 to 4 times more efficient than existing open models. Beam offers a high-performance Western open-weight alternative to leading Chinese open models like GLM-5.2. Its high computational efficiency substantially lowers the cost of running advanced coding assistants and autonomous AI agents. Beam uses a sparse Mixture-of-Experts (MoE) architecture with 501 billion total parameters, but only activates 23 billion parameters per token. Reflection AI, currently valued at $25 billion, is already training its next-generation model.
## BACKGROUND
A Mixture-of-Experts (MoE) model routes inputs to specific specialized subnetworks (experts) rather than using the entire network for every calculation, making large models much faster and cheaper to run. Agentic AI refers to systems capable of autonomous planning, tool usage, and multi-step task completion with minimal human supervision.