Reflection releases open model Beam
What happened
On October 6, 2026, Latent Space reported that Reflection announced Beam. It is a text-only mixture-of-experts model trained from scratch, with 501 billion total parameters and 23 billion active per pass, aimed at coding, agent and science tasks. The team also disclosed pretraining used 23.8 trillion tokens. Beam has been announced, and the full weights are planned for release this month under an Apache 2.0 license; the report does not say the full weights are already public.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Latent Space[AINews] Reflection Beam - 501B-A23B American Open Model
Reflection 宣布推出 Beam，这是一款从零训练的纯文本 MoE，拥有 501B 总参数和 23B 激活参数，面向编码、智能体及科学任务。完整权重计划本月以 Apache 2.0 许可发布，团队披露预训练使用了 23.8T 模型 tokens。
Heat over time
Not enough continuous observations to draw a trend yet.