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

Reflection Releases 501B Open Model for On-Premise AI

The new US-based AI lab Reflection has released Beam, a 501-billion parameter Mixture-of-Experts (MoE) model, making it available on Hugging Face under a permissive Apache 2.0 license. The model is designed for teams seeking to run and customize a frontier-scale model on their own hardware. With a 1 million token context window and a specific focus on coding and agentic workflows, Beam provides a new option for organizations that require data privacy, custom fine-tuning, and direct control over their AI infrastructure.

Beam's architecture uses a sparse MoE design, which keeps the computational cost relatively low by activating only 23 billion of its 501 billion parameters for any given token. This allows it to run at the speed of a much smaller model while retaining the capacity of its larger size. Key technical specifications include:

• Architecture: Sparse Mixture-of-Experts (MoE) with 256 experts per MoE layer.
• Attention: A hybrid system using sliding-window attention for most layers and global attention for every sixth layer to efficiently manage its 1M token context.
• Performance: The model ships with a custom Rust-based harness and has day-zero support in libraries like vLLM and SGLang.
• Hardware: Practical deployment requires significant memory, with the NVFP4-quantized checkpoint needing ~280GB. This makes it suitable for multi-GPU setups using hardware like the NVIDIA B200.

The availability of Beam signals a continued push towards powerful, open models that can directly compete with proprietary systems. By providing recipes for supervised fine-tuning (SFT) and reinforcement learning (RL), Reflection is encouraging the development of a specialized ecosystem around its model. This move empowers companies with the necessary hardware and technical expertise to build highly customized, long-context agents and coding assistants without relying on third-party APIs, potentially shifting a segment of the market away from closed, centralized AI providers.

Reflection's release of Beam is a direct challenge to the closed-source model ecosystem, offering enterprises a path to sovereign, customizable AI at scale, provided they have the substantial hardware to support it.