The mixture-of-experts model contains 501B total parameters and 23B active parameters with a 1M token context window. Trained on Nvidia GB300 NVL72 hardware, Beam delivers over 70K tokens per second per GPU on Blackwell Ultra. It has day-0 support from vLLM, SGLang, and Ollama, with the latter providing the model for free for a limited time. Artificial Analysis ranks Beam as the second-highest US open weights model on its Intelligence Index and the leading US open model on its Cyber Index.
Tuesday, October 20, 2026
Story
Reflection AI Releases 501B Beam Weights as First Open Model
AI5 days ago
Reflection AI has published the weights, tokenizer, and reference implementation for Beam under an Apache 2.0 license, making the model available on Hugging Face. The release includes Mirror CLI, the open-source training harness Reflection used to build the model in collaboration with more than 40 partners. Designed for software engineering and long-horizon agentic tasks, the model is now available for deployment across multiple platforms.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.