TL;DR of the paper, generated automatically.

NVIDIA's 4B Dory1.0 loops one shared block to trade compute for accuracy.

• Looping: more loops improve accuracy, most on coding, agentic and long-context tasks.
• Efficiency: four loops give 1.3-2.5x fewer tokens than one, on agentic benchmarks.
• Release: open checkpoints in BF16, MXFP8, NVFP4 and GGUF.