Hugging Face introduces Reflection Beam, a 501B open model

Hugging Face has introduced Beam, a new open model from the US lab Reflection. The announcement, published on October 16, 2026, describes Beam as a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters per token. It accepts text, offers a context window of one million tokens and is released under the Apache 2.0 license. Hugging Face says the model is aimed at coding and agentic work rather than general multimodal use.

The architecture explains the difference between its headline size and its per-token compute. Beam has 256 experts in each mixture-of-experts layer, but routes each token through eight of them. All experts still need to remain in memory. Hugging Face says the BF16 checkpoint is about one terabyte in safetensors format, while the NVFP4 version is about 280 GB and is designed for NVIDIA Blackwell GPUs. FP8 checkpoints are also available for older Hopper and Ampere hardware. In other words, the model is computationally closer to a smaller model for each token, but it still requires unusually large infrastructure to load.

The release includes a Rust-based custom harness from Reflection and day-one support in Transformers, vLLM and SGLang. Hugging Face also provides recipes for supervised fine-tuning and reinforcement learning, including LoRA adaptations that keep the base weights frozen. The model can be served on dedicated GPU systems or tested through Hugging Face Inference Providers, with Baseten, Fireworks and Together listed in the announcement.

The practical significance is portability. Teams that want to experiment with a frontier-scale model do not have to rely only on a single hosted assistant: they can inspect the weights, adapt the model and choose their serving stack. That freedom comes with a high hardware bill. Hugging Face recommends several B200 or B300 GPUs for practical serving, especially when long contexts or concurrent requests are involved. The published evaluations are described as “vibe evals,” not formal benchmarks, so they are useful examples rather than an independent ranking. For makers and companies, Beam is therefore an important open release, but its real value will depend on reproducible testing, operating cost, license compliance and safe controls around agent actions.