Reflection's Beam becomes the most capable open-weight model built outside China
decoder ·
AI startup Reflection has announced Beam, its first open-weight model designed for coding and reasoning with a focus on compute efficiency. The mixture-of-experts model contains 501 billion parameters but activates only 23 billion per token, allowing it to match the performance of GLM 5.2 while using three to four times less compute. Trained over four weeks using 10,500 Nvidia GPUs, Beam demonstrates strong capabilities in agentic tasks and software engineering benchmarks. Notably, the model exhibited emergent abilities such as web browsing during reinforcement learning, despite not being explicitly trained for those tasks. Reflection also developed a separate safety and alignment model to ensure factual accuracy and adherence to hard rules. The company, founded by former DeepMind researchers and backed by significant funding, positions Beam as a cost-effective Western alternative to Chinese models like Qwen and DeepSeek. Beam will be released under the Apache 2.0 license later this month, following final safety testing.
decoder