Frontier Open Weights
Reflection AI unveils Beam, a 501B open-weight model trained on 10,500 GPUs
Reflection AI introduced Beam, a sparse 501B-parameter mixture-of-experts model with only 23B active parameters, delivering 3-4x better inference efficiency than comparable frontier models. Trained on a cluster of 10,500 NVIDIA GB300 GPUs running 46.4 million sandboxes per day for 28 days, Beam advances coding, reasoning, and agentic task performance with full open-weight release under Apache 2.0 expected later this month.
