23 October 2026

Reflection AI releases open weights for its Beam model

First reported

Deep Learning Weekly ran this on .

  • Reflection AI, an AI company, has published the trained weights (the learned settings that make a model work) for its model Beam.
  • Beam has roughly 500 billion parameters (the adjustable numbers inside a model), a very large size.
  • Its training ran for about two months across several thousand graphics processors, with almost no downtime.

How it was covered

Deep Learning WeeklyEditorial team

Model scale, training run and cluster reliability, and released weights and recipes