Reflection 发布 501B-A23B 美国开源模型 Beam
原标题：[AINews] Reflection Beam - 501B-A23B American Open Model
AI 摘要
Reflection 发布 Beam，一个 501B 总参数、23B 激活的 MoE 文本模型，面向编程、智能体与科研任务，从零训练，Apache 2.0 全权重将于本月发布。团队称使用 23.8T 预训练 token，并在约 10,500 块 GB300 上完成预训练与 RL，SWE-bench Verified 得分 80.9。第三方分析认为其智能水平大致与 GLM-5.2 相当，但落后于 DeepSeek V4 Flash 等中国模型，同时被视为美国开源模型生态的重要新成员。
正文节选
It’s been over a year since Reflection launched with us with big goals on coding (and hinted about their RL approach): But they stayed “stealth” longer than Thinking Machines and it was not clear we would ever get a model launch out of them, as the broader open model ecosystem did not slow down one bigt for them. Well, we did: They compare themselves to Inkling, Nemotron, and GLM 5.2, but the SOTA GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash are generally ahead. Still, since this is U