• Sources: primary, discussion
  • Summary: Reflection's announcement post describes Beam as a mixture-of-experts model of 501B total and 23B active parameters across 52 layers, pretrained on 23.8T tokens, with a reinforcement learning run of more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks at up to 256K context. Nothing is downloadable: the post states Beam is in final red-teaming and that weights, a technical report, a model card and developer artifacts arrive later this month, with no date given. Every benchmark figure is Reflection's own, comparison scores are taken from Artificial Analysis and DataCurve rather than reproduced, and the claim of comparable reasoning scores at 3 to 4 times less inference compute than GLM-5.2 rests on an estimated-FLOPs model the post itself calls approximate.
  • Why it matters: The parameter shape and the training run are the only checkable parts of this announcement, because the weights, the report and the model card that would let anyone verify the rest do not exist publicly yet.
  • Follow-up: Watch for the weight release and its license, the technical report and model card, and any independent benchmark or inference-cost reproduction.