Welcome Beam: a 501B open model you can own, customize, and scale

Reflection has released Beam, a 501B parameter open-source Mixture-of-Experts (MoE) model designed for coding and agentic tasks. The model features a 1 million token context window, 23B active parameters per token, and is available under an Apache 2.0 license. Its architecture includes a unique residual stream weighting, a hybrid attention mechanism with sliding-window and global attention, and a group-limited expert routing scheme. Support for inference is provided through libraries like transformers, vLLM, and SGLang, with recipes available for supervised fine-tuning and reinforcement learning.