Beam 1.0

Beam 1.0 is a large-scale open-weight language model developed by Reflection. Built as a sparse mixture-of-experts (MoE) architecture, the model features a total parameter count of 501 billion with 23 billion active parameters per token. It supports an extensive 1 million token context window, making it suitable for long-horizon workflows and comprehensive data analysis.

The architecture comprises 52 layers—utilizing 3 dense layers and 49 MoE layers—with 256 experts per MoE layer and 8 routed experts per token. It incorporates gated grouped-query attention combined with sliding window attention. Beam 1.0 was pretrained on 23.8 trillion curated tokens covering diverse web text, source code, and technical documentation, followed by intensive reinforcement learning optimization designed to enhance efficiency and capabilities.

Capabilities and Reasoning

Beam 1.0 is engineered with a primary focus on software engineering, agentic tool use, and complex reasoning. The model achieves high token efficiency, executing advanced multi-step workflows, resolving real-world programming issues, and supporting tool integration via Model Context Protocol servers. Its reasoning capabilities extend to graduate-level mathematics and scientific problem solving, maintaining contextual consistency across extended interaction turns.