Dory-Beam-4B vs base Dory, side by side

NVIDIA's looped Dory1.0-4B, distilled from Reflection's Beam-1.0-501B-A23B on 52,660 verified Beam traces and blended back into the base weights. It reasons about half as long as base Dory at nearly the same accuracy. Below are real answers from both models to the same evaluation problems.

Model GGUF for llama.cpp Beam traces dataset Base model

Real time, on a laptop

Both models answering the same MATH-500 problem with llama.cpp on an Apple M4 Max (Q5_K_M, 2 loops, one run each). Recorded token timings, played back in real time. The problem was chosen before recording as the one with the median token ratio in the full eval.

Six panels, one per benchmark, plotting accuracy against mean output tokens for base Dory and Dory-Beam-4B at 1, 2 and 4 loops. Dory-Beam's points sit at similar accuracy but roughly half the tokens.
Base DoryDory-Beam-4B 2 loops, every sample in the full evaluation
BenchmarkBase accuracyDory-Beam accuracy Base tokensDory-Beam tokensFewer tokens