Nvidia-backed Reflection unveils Beam, a 501-billion-parameter open-weight AI model
Reflection AI says Beam can match Z.ai’s GLM-5.2 on advanced reasoning while using three to four times less inference compute; it is designed for coding and AI-agent tasks. The model is still undergoing safety evaluations, with its weights and technical details due later this month, so European businesses cannot yet assess its performance for themselves.
Bottom line — Beam activates 23 billion of its 501 billion parameters per task, but independent testing must wait for its planned release.
Go deeper 6
-
Reflection says Beam approaches Alibaba’s Qwen 3.8-Max on coding and agentic tasks, while its own comparisons put Moonshot AI’s Kimi K3 ahead on raw capability.
-
The company’s reported benchmark table gives Beam 80.1 on Terminal Bench v2.1, against 88.3 for Kimi K3; The Next Web notes the figures have not been independently verified.
-
Reflection says the three-to-four-times efficiency comparison is an approximate inference-compute estimate, not a measured operating-cost saving.
-
According to Reflection, pretraining used 6,144 Nvidia GB300 GPUs and 23.8 trillion tokens; its four-week reinforcement-learning run used 10,500 GPUs and generated more than 100 million rollouts.
-
Reflection says Beam is in final red-teaming and evaluations, with early access limited to selected users. The planned release includes weights, a technical report, a model card and developer tools under an Apache 2.0 licence.
-
The Next Web reports that the UK’s AI Safety Institute and a US government body are assessing Beam with Reflection; the company says its safety results will be published in the technical report.