Reflection 发布 MoE 开源模型 Beam，总参数 501B、激活 23B，从零端到端训练，聚焦编码、推理和智能体任务，本月发布全部权重。官方称推理效率比 GLM 5.2 高 3–4 倍，比同类领先西方开源模型高 4 倍以上；预训练用 24T token 耗时四周，随后在 10,500 张 GB300 上做了四周强化学习，SWE-bench Verified 得分 80.9。
Reflection has announced Beam: a 501B-parameter model with 23B active, with full weights scheduled for release this month. Frontier capabilities, and holy is this model efficient!
Trained from scratch, Beam targets coding, reasoning and agentic tasks.
Reflection claims 3–4x higher inference efficiency than GLM 5.2 and over 4x higher efficiency than leading Western open models in its class.
24T pretraining tokens in four weeks, followed by four weeks of reinforcement learning on 10,500 GB300s.
Terminal Bench v2.1 and 80.9 on SWE-bench Verified.
But again: look at the efficency, eypecially in Terminal bench v2.1! Congrats on that launch. More western open models to come.
Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active. - Frontier reasoning efficiency - Advances the Western open frontier on coding & agentic tasks - Trained end-to-end from scratch Full weights release this month. Learn more about Beam: http://reflection.ai/beam在 X 查看被引用的帖子
来源：Chubby♨️ · x.com