Reflection AI on October 16, 2026, released the weights for its first open-source model, Beam, under a permissive Apache 2.0 license. The 501-billion-parameter model was trained using reinforcement learning that penalizes the use of unnecessary tokens, a technique designed to make it more efficient and cost-effective for completing tasks compared to rivals.
Beam is positioned as a U.S. competitor to the dominant open-source models from Chinese labs. Founded by former Google DeepMind researchers and backed by investors including Nvidia, Reflection AI states that Beam is designed for enterprise coding and complex reasoning tasks. While the company claims Beam can match the performance of some leading Chinese models with significantly less computing power, its performance has not yet been independently verified.