Reflection's Beam aims to undercut Chinese open models with fewer tokens, not cheaper tokens

Reflection released Beam, a 501-billion-parameter open model trained to finish tasks with fewer tokens rather than cheaper ones. During reinforcement learning it was penalized for using unnecessary tokens, and the company says a naive penalty would teach a model to give up, so it ran a very large training effort. Fewer tokens per finished task can cut the bill even when each token costs more.

Summary written with AI by HowTechMoves, based on reporting by VentureBeat.