Beam, a new 501-billion-parameter open-source language model from Reflection—a newly launched US AI lab—has arrived on Hugging Face. The mixture-of-experts model, released under the Apache 2.0 license on October 16, represents a shift toward open, enterprise-scale AI infrastructure. With only 23 billion active parameters per token, Beam delivers performance comparable to much larger dense models while remaining deployable on private infrastructure, cloud systems, or customer hardware. The model is engineered for reasoning-intensive and agentic applications, particularly in coding. It features a 1-million-token context window, native support across transformers, SGLang, and vLLM inference libraries, and is backed by a custom Rust-based inference harness for optimal performance. Reflection has bundled supervised fine-tuning and reinforcement learning recipes alongside the base model, enabling teams to customize Beam for domain-specific tasks. Technical innovations include a scaled-softmax query mechanism for improved attention precision at long context lengths and a group-limited expert router inspired by DeepSeek-V3. To catalyze developer adoption, Reflection is distributing $10,000 in Hugging Face Hub credits to creators of the most impactful projects built with Beam—whether demos, quantizations, synthetic datasets, or benchmarks. The open release underscores a growing market segment: frontier-scale models designed not for API consumption but for organizations seeking full control over their inference infrastructure and fine-tuning pipelines.