Reflection AI Releases Beam: A 501B Parameter Open-Source MoE Model for Coding and Agents

AI startup Reflection released its first open-source large language model, Beam (Beam-1.0-501B-A23B), on October 16. The Mixture-of-Experts (MoE) model has 501 billion total parameters but activates only about 23 billion per token, aiming to balance the capacity of a large model with the inference speed of a smaller one. Released under the Apache 2.0 license, it supports text input/output and features a 1 million token context window.

Beam is targeted at coding and agentic use cases. Official checkpoints are available in BF16, FP8, and NVFP4 formats, with the NVFP4 version optimized for NVIDIA Blackwell GPUs, reducing memory requirements from approximately 1 TB to around 280 GB. According to the Hugging Face blog, the model has received 'Day-0' support from major inference libraries like transformers, vLLM, and SGLang, along with recipes for supervised fine-tuning (SFT) and reinforcement learning (RL) to facilitate customization on private infrastructure.

Based on benchmark results provided by the Reflection team, Beam scores 43.3% on DeepSWE v1.1 and 76.5% on SWE Bench Pro v2-Hard (at BF16 precision), achieving 97.7% on the AIME 2026 math reasoning test. These performance figures are currently self-reported by the vendor and have not yet been independently reproduced.