Beam is a sparse expert-based AI model with 501 billion parameters, efficient for coding and reasoning tasks(reflection.ai)
Writing language: KoreanRead in the original language
Plain summary
- Beam, developed by Reflection, is a sparse mixture-of-experts (MoE) model with 501 billion parameters, of which 23 billion are activated, designed specifically for coding, reasoning, and agent tasks.
- The model was pretrained on 23.8 trillion high-quality web and licensed data tokens and trained using over 100 million reinforcement learning rollouts by running 10,500 NVIDIA GB300 GPUs continuously for 4 weeks.
- Reflection stated that Beam shows competitive performance on coding and agent-related tasks compared to public models of the same or larger scale like GLM 5.2 and Qwen 3.8-Max, while achieving greatly improved efficiency by using 3 to 4 times less GPU compute during inference.
- The training employed a newly optimized asynchronous policy gradient algorithm for reinforcement learning and utilized about 1 million task environments. It also applied various techniques for balanced expert utilization and stable signal propagation, ensuring overall training stability.
Found on
Hacker News ↗ / 462 votes / 148 comments
Have you run into something like this? Tell us in a line or two what you did. You do not need to know the product.