TLDR Tech / Launches
Reflection's Beam, an efficient open-weight model, leads outside China
AI firm Reflection has unveiled Beam, its initial open-weight model designed for coding and reasoning tasks with an emphasis on computational efficiency. This mixture-of-experts model utilizes 23 billion of its 501 billion parameters per token, achieving parity with GLM 5.2 on key benchmarks while consuming substantially less processing power. Beam, trained using reinforcement learning on 10,500 Nvidia GPUs, is scheduled for release under the Apache 2.0 license this month.
Read the full story