Reflection releases Beam’s open weights, betting on fewer tokens per task

The coding-focused model carries an Apache 2.0 license. Its efficiency pitch rests on company estimates that exclude prompt-reading work—not measured deployment savings.