Hugging Face introduces Beam, a 501B open model with substantial hosting requirements

Hugging Face says Reflection’s Beam is a text-only, Apache 2.0 mixture-of-experts model with 501B parameters, 23B active per token and a 1M-token context window, aimed at coding and agentic work. That offers teams a self-hosting and customization option. Source

Hugging Face cautions that all experts remain in memory: sparse computation does not remove the full model’s storage burden. Its approximately 280GB NVFP4 checkpoint needs Blackwell hardware for native execution; Hugging Face recommends four B200 GPUs for long contexts and concurrent requests. Deployment capacity therefore remains a significant consideration. Source

Hugging Face attributes benchmark results to Reflection and describes its own five-task checks as informal evaluations, not benchmarks. Those results do not establish production reliability, workload-specific costs or sustained throughput; these remain unverified. Source