Reflection AI’s Beam: a 501B open model for code and agents
Reflection AI introduced Beam for coding and agentic work, and promises to publish the weights under Apache 2.0.
Reflection AI has introduced Beam, a workhorse model for enterprise coding and agentic tasks. The main promise is open weights under Apache 2.0, which the company plans to publish this month.
What’s new
Beam is a Mixture-of-Experts model with 501B total parameters and 23B active. It’s text-only. Reflection pretrained it on 23.8 trillion tokens and added reinforcement learning: over 100 million rollouts on 10.5K NVIDIA GB300 GPUs across 4 weeks.
By Reflection’s own numbers, Beam scores 80.9 on SWE-Bench Verified and 80.1 on Terminal-Bench 2.1. The company claims it needs 3–4x less inference compute than GLM-5.2 on advanced reasoning benchmarks at comparable scores.
How to try it
Early access and a waitlist are on platform.reflection.ai, with docs on developers.reflection.ai. The weights are due in the company’s Hugging Face collection. Reasoning effort is adjustable, trading response length against performance.
Limitations
At announcement, the model was still going through final safety evaluations. Pricing isn’t stated, and all benchmark results come from the company, so independent testing is worth waiting for.
Prepared with AI from public sources.