Reflection Unveils Beam, A 501B Open-Weight Model Aimed At Chinese Open-Source AI
Reflection is pitching Beam's inference efficiency, not raw capability, as the lever against larger Chinese open models, and the weights are promised for October 2026.
Reporting from 1 source: GIGAZINE.
Reflection announced Beam, an open model with 501 billion parameters and 23 billion active parameters. It was pretrained on 23.8 trillion tokens and trained with reinforcement learning across 10,500 NVIDIA GB300s for four weeks, generating over 100 million rollouts. Reflection says Beam matches GLM 5.2 on coding and agentic tasks at one-third to one-quarter the inference compute, and approaches Qwen 3.8-Max.
Beam is not out yet. Reflection says the model is still in final-stage red teaming and evaluation, and that it plans to release the full weights in October 2026, with a waiting list open through its playground.
The comparison chart puts Beam against GLM 5.2, Qwen 3.8-Max, Inkling, and Nemotron Ultra across DeepSWE, Terminal Bench, HLE No Tools, SWE Bench Pro, SWE Bench Verified, and CritPT AA. Reflection's own claim is parity on those benchmarks, with the efficiency gap widening against models above 2 trillion parameters.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.