Reflection AI introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, aimed at coding and agentic work. The company says it pretrained on 23.8 trillion tokens and ran more than 100 million reinforcement-learning rollouts on 10,500 NVIDIA GB300 GPUs.
The systems disclosure is more interesting than the leaderboard copy: 23 billion active parameters suggests an aggressive inference-efficiency bet, and Reflection says it built mechanisms to update weights across the fleet quickly while recovering from inference failures during training.
But keep the status exact. Beam is in final red-teaming and evaluation, available only to select early-access users. Reflection says weights, a technical report, model card and developer artifacts will arrive later in October under Apache 2.0. Until then, its benchmark comparisons remain vendor claims rather than independently inspectable results. Reflection AI