Reflection Releases Beam-1.0-501B-A23B Open MoE LLM
Reflection has released Beam-1.0-501B-A23B, a sparse mixture-of-experts open LLM with 501.3B total parameters, 23.4B active parameters, and a 1M token context window. The model is available in BF16, FP8, and NVFP4 formats with specialized capabilities in software engineering and token-efficient reasoning. This release marks a major advancement for the open-weights AI ecosystem by offering massive scale combined with strong performance in agentic tool use and software engineering tasks. Its ability to achieve high solution quality using fewer inference-time reasoning tokens addresses a critical efficiency bottleneck in large language models. Beam features 52 layers with Gated GQA and Sliding Window Attention, a 152k token vocabulary, and supports advanced agentic workflows including Model Context Protocol (MCP) tool integration. It achieves strong benchmark scores such as 81.6% on SWE-Bench Verified and 90.6% on GPQA Diamond.
## BACKGROUND
A sparse mixture-of-experts (MoE) architecture routes inputs to a subset of neural network parameters, allowing models to scale up total capacity while keeping active compute costs manageable. NVFP4 is a low-precision 4-bit floating-point format designed to accelerate inference on modern hardware architectures, while MCP standardizes how AI agents securely communicate with external tools and data sources.