Reflection releases Beam, an open 501-billion-parameter coding model under the Apache 2.0 license
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A large model for writing code that you can download and run yourself has become more accessible: open weights are narrowing the gap with closed systems.
Beam is a sparse mixture-of-experts model with 501 billion parameters, 23 billion active per token and a 1-million-token context; its weights are on Hugging Face in BF16, FP8 and NVFP4 formats. According to its authors, it scores about 78.5% on SWE-Bench Pro v2-Hard, 78-80% on Terminal Bench v2.1 and 97.5-98% on AIME 2026. Running it takes at least four B200 GPUs or eight H200s. The model is available from inference providers (Baseten, Fireworks, Together), through vLLM and SGLang, and on AWS in SageMaker JumpStart and HyperPod.
Sources 2