Reflection previews 501B MoE model Beam with 23B active parameters and 100M+ RL rollouts, with weights planned this month

What changed

Reflection AI introduced Beam on October 5, a sparse MoE with 501B total parameters and 23B active parameters aimed at coding, reasoning and agentic workloads. Reflection says pretraining covered 23.8T tokens, while a four-week high-compute RL run on roughly 10.5K NVIDIA GB300 GPUs generated more than 100 million rollouts; midtraining extended effective context to one million tokens. Beam is still undergoing final red-teaming and evaluation and is available only to a select early-access group. Reflection says Apache 2.0 weights, the technical report, model card and developer artifacts will be released later this month, so the current announcement should not be described as downloadable open weights already being generally available.