Reflection AI has introduced Beam, a 501-billion-parameter open-weight model for coding and agent tasks, and plans to release it under an Apache 2.0 license later this month. The model uses a mixture-of-experts architecture, was trained with large-scale reinforcement learning, and is still undergoing safety evaluation and red-teaming before broader access. #ReflectionAI #Beam #Apache20
Keypoints
- Beam is a 501-billion-parameter open-weight model for coding and agent tasks.
- Reflection plans to publish Beam’s weights under an Apache 2.0 license.
- The model uses mixture-of-experts architecture to reduce inference cost.
- Beam trails top open models on some benchmarks but uses less compute per answer.
- Reflection is still completing red-teaming, safety evaluation, and a technical report.