Beam: Reflection’s 501B‑Parameter Open Mixture‑of‑Experts Model

Reflection lab has released Beam, a 501‑billion‑parameter mixture‑of‑experts language model that activates only 23 billion weights per token.

The model uses a sparse MoE design with 52 transformer layers, 256 experts per layer and routes eight experts for each input token. It supports a one‑million‑token context window, accepts text‑only input and is distributed under the Apache 2.0 license.

Why it matters for GPU / AI infrastructure

Because the active parameter count stays low, Beam can run on existing GPU clusters with lower memory footprint while still delivering the representational power of a half‑trillion‑parameter model. This makes it attractive for enterprises that want to host frontier‑scale LLMs on‑premise or in private clouds without investing in exotic hardware.

Beam ships with an open‑source Rust inference harness and day‑0 support in 🤗 Transformers, vLLM and SGLang, plus ready‑to‑use recipes for supervised fine‑tuning and reinforcement learning. Reflection is also offering $10,000 in Hugging Face Hub credits to the community for the best projects built with Beam.