Reflection开源501B MoE编程推理模型
推荐指数 71.0 NO. 009 · 2026.10.06
发布2026/10/05Score262Comments66
为什么值得看
Reflection发布Beam，501B总参数、23B激活参数的稀疏MoE模型，针对代码、推理和agent场景预训练了23.8T token，并通过大规模RL后训练。对做AI编程工具或需要本地部署强推理能力的团队来说，多了一个可直接商用的开源基座选择。
编辑判断
Beam的23.8T token预训练量远超Llama 3 70B的15T，但501B总参数只有23B激活，说明Reflection在走"大力出奇迹+推理降本"的双轨路线。RL阶段10500块GPU跑出1亿次rollout，这个规模接近OpenAI内部训练强度，不是普通开源团队能复制的。
真正值得盯的是它的agent能力定位——目前开源社区缺乏能跟Claude 3.5 Sonnet在工具调用上正面较量的模型。如果Beam的RL确实针对多步工具使用做了优化，Cursor、Devin这类产品的底层模型选择会多一个选项。建议直接跑它的SWE-bench和LiveCodeBench成绩，比官方宣传更有说服力。
社区反馈
意见分歧 62 条评论
核心争论：开源模型权重未发布遭质疑，数据保密成行业常态
> Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads. > Beam’s capabilities come from major investments in both pretraining and reinforcement learning (RL). We pretrained the model on 23.8 trillion divers
And also a "proprietary data set" hahaha... Probably just means they don't want to show it, and it is data, that either they shouldn't have, or that there is nothing special about their training data and it is just meant to sound like there is some secret ingredient, while there is none.
Not sharing the data is pretty standard because 1) it tends to get the lawyers involved and 2) good data is critical for getting good results. Imo you can get better results with great data and generic modeling techniques than with incredible modeling techniques and crappy data. Because if you have