[Reflection Open-Sources Beam Weights, 501B Parameters MoE Using Apache 2.0] DeepFlow Tech News: On October 16, Reflection released its first open-source weight model, Beam, with weights made available on Hugging Face under the Apache 2.0 license. Beam is a sparse MoE model with a total of 501B parameters and 23B activations, designed for encoding, inference, and long-range agent tasks. It provides BF16, FP8, NVFP4 precision weights and speculative decoding drafter models. The release of this model had been previously reported, and this update includes new details on the open-source license and weight precision.