The model is a mixture-of-experts design, meaning only 23 billion of its 501 billion total parameters activate per token, which keeps running costs down. Reflection says Beam matches GLM 5.2 on key reasoning benchmarks while using three to four times less compute, and comes close to the larger Qwen3.8-Max on coding and agent benchmarks. The efficiency claim comes with a caveat Reflection itself makes: stronger open models such as Kimi K3 still beat Beam on raw performance, and the company says a successor is already in training to close that gap.

Beam's abilities come from large-scale pretraining plus a heavy reinforcement learning phase, which Reflection says ran 10,500 Nvidia GB300 GPUs for over four weeks, one of the largest training runs any open lab has done. Scores kept climbing throughout, with no plateau even past 80 million rollouts. Reflection also reports an unexpected side effect: with no browsing tasks in its training mix, Beam improved at web browsing anyway, learning to query other language models and pull documents from external services.

A tunable parameter lets developers choose between fast answers and longer reasoning on harder tasks, trading compute cost against quality. For safety, Reflection trained a separate model and merged it with Beam, and plans to publish safety test results and open-source its evaluation methods. Beam is text-only, though it can process other media when represented as text.