Researchers introduced Dory1.0-4B, a 4-billion-parameter language model whose shared block can run repeatedly at inference. That lets one checkpoint use different amounts of computation, rather than requiring a separate larger model for each budget.
The paper reports that using two or more loops beat several larger models on average across 12 benchmarks. With four loops, Dory generated 1.3–2.5 times fewer tokens on agentic benchmarks than with one loop. The researchers say the model’s KV-cache footprint stays the same at any loop count, and they have released checkpoints in four formats.
Why it matters
For developers choosing an open model, Dory offers a way to spend more inference computation without increasing its KV-cache footprint, though the reported benchmark gains are from the paper and do not establish results in every application.
