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

Reflection, a new US AI lab, has released its first model, Beam, on Hugging Face under the Apache 2.0 license. In a post dated October 16, 2026, Hugging Face describes a text-only mixture-of-experts system with 501 billion parameters, of which about 23 billion are active per token, and a context window of 1 million tokens. It is aimed at coding and agentic work for teams that want a frontier-scale model on their own infrastructure.

Each of 49 expert layers holds 256 experts and routes each token to 8, using a scheme modeled on DeepSeek-V3. Hugging Face says this gives the speed of a 23B model, though all weights must stay in memory. The full-precision checkpoint is about 1 TB, while the NVFP4 version shrinks to roughly 280 GB. Hugging Face recommends planning for four Nvidia B200 GPUs.

Benchmarks supplied by Reflection include 81.6% on SWEBench Verified, 80.1% on Terminal Bench v2.1, 97.7% on AIME 2026 and 43.3% on DeepSWE v1.1. Hugging Face's own informal tests passed all five tasks at high reasoning effort, but it cautions these are not benchmarks.

Beam launched with support in transformers, vLLM and SGLang, availability through Baseten, Fireworks and Together, plus fine-tuning and reinforcement learning recipes. Reflection is also offering $10,000 in Hub credits for top community projects. The release matters as a large, permissively licensed model, though one commenter noted its hardware costs could put it out of reach for all but large organizations.