Reflection's Beam becomes the most capable open-weight model built outside China
Reflection's Beam becomes the most capable open-weight model built outside China
Reflection推出了其首款开放权重模型Beam。该专家混合系统每标记仅激活5010亿个参数中的230亿个，旨在与GLM 5.2在编码和推理方面相媲美，同时使用的计算量仅为后者的三到四分之一。这种策略是针对中国开放权重竞争对手如Deepseek和Qwen，强调效率而非单纯性能。文章《Reflection的Beam成为在中国以外打造的最强大开放权重模型》最先出现在The Decoder上。
Reflection has released Beam, its first open-weight model. The mixture-of-experts system activates just 23 billion of its 501 billion parameters per token and aims to match GLM 5.2 on coding and reasoning while using three to four times less compute. The strategy against Chinese open-weight rivals like Deepseek and Qwen is efficiency over raw performance. The article Reflection's Beam becomes the most capable open-weight model built outside China appeared first on The Decoder .
应来源方要求，这里只提供摘要与原文入口。完整内容请阅读原文。
来源：The Decoder · the-decoder.com