Reflection releases Beam, a 501B-parameter sparse MoE model
What happened
On October 6, Hacker News' front page covered Reflection's sparse mixture-of-experts model Beam, aimed at coding, reasoning and AI agent tasks, with 501 billion total parameters and 23 billion active, and weights planned to open this month. Later the same day, TechCrunch reported that Reflection AI had officially released Beam, a text-only open-weight model with a 1 million token context window. The company claims its advanced reasoning matches GLM-5.2 at 3 to 4 times less inference compute, but those performance claims have not been independently verified. Earlier coverage said weights would open this month; later coverage said the model is officially released, without stating whether the weights are downloadable yet.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- TechCrunch · AIReflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost
Reflection AI 正式发布纯文本 MoE 模型 Beam，声称其高级推理表现与 GLM-5.2 相当，推理计算量则低 3-4 倍。Beam 拥有 501 billion 总参数、23 billion 活跃参数和 1 million token 上下文窗口，面向推理、编码与智能体任务，上述性能主张尚未经过独立验证。
- Hacker News front pageBeam: Reflection's 501B open-weight model
Reflection 介绍面向编码、推理和 AI 智能体任务的稀疏 MoE 模型 Beam，总参数为 501 billion，激活参数为 23 billion。
Heat over time
Not enough continuous observations to draw a trend yet.