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

Hugging Face has released Beam, an open mixture-of-experts (MoE) model developed by the new US lab Reflection, featuring 501 billion parameters with 23 billion active per token. The model supports text input, operates with a context window of 1 million tokens, and is licensed under Apache 2.0. Beam is optimized for coding and agentic applications, targeting teams seeking frontier-scale customization on their own hardware or cloud infrastructure.

Beam includes a Rust-based open-source harness for performance optimization, alongside native support in Transformers, vLLM, and SGLang libraries. Hugging Face and Reflection provide inference, supervised fine-tuning, and reinforcement learning recipes to facilitate post-training and deployment. The model employs a sparse MoE architecture where only a fraction of the network processes each token, enabling efficient inference at the speed of a smaller model while maintaining large-scale capacity.

The model’s architecture features 52 decoder layers with a hidden size of 5,120, using a combination of sliding-window and global attention mechanisms. Sliding-window layers attend to the previous 512 tokens, while global layers attend to the full context, implemented with scalable-softmax query scaling to maintain precision. The MoE block contains 256 experts, with only 8 activated per token, and uses a group-limited routing scheme inspired by DeepSeek-V3.

Hugging Face is offering $10,000 in Hub credits to projects that build and share applications using Beam. Developers can deploy the model via vLLM or SGLang recipes, with checkpoints available in BF16, NVFP4 for Blackwell GPUs, and FP8 for older hardware. Practical deployment requires significant GPU memory, such as four NVIDIA B200 GPUs (720 GB total) for optimal performance with long contexts and concurrent requests.