Reflection Unveils Beam, a New 501B Open-Weight AI Model

Reflection has introduced Beam, a substantial 501B-parameter sparse mixture-of-experts AI model with 23B active parameters, trained on 23.8T tokens using 10.5K Nvidia GB300 GPUs over four weeks. The weights, technical report, and developer artifacts are expected to be released later this month, offering a powerful open-weight alternative in the AI landscape.