NVIDIA has released Nemotron Labs Dory1.0-4B, a 4-billion-parameter looped language model whose inference depth developers can adjust at runtime, posting model cards for BF16, NVFP4 and MXFP8 variants on Hugging Face under an OpenMDW license. The model supports context lengths up to 256K tokens.
Dory divides its layer stack into input, core and output blocks and reinjects the input into every loop of a shared core block. Because the same parameters are applied repeatedly, users can trade inference compute for accuracy per task without retraining: NVIDIA recommends two loops as a default starting point and four loops when accuracy matters more than speed.
On NVIDIA's reported benchmarks, Dory1.0-4B scores 50.3 on SWE-bench Verified at one loop, 59.5 at two loops and 61.8 at four loops - compared with 54.7 for Qwen3.5-4B, 61 for Qwen3.5-9B and 57.40 for Gemma 4 26B A4B on the same table. Terminal-Bench 2.1 scores run 12.5, 18.1 and 22.2 across the same loop counts.
The architecture targets memory-constrained local deployment: Dory supports last-loop KV caching, which avoids storing a separate KV cache for every loop, and the card describes native NVFP4 pre-training. NVIDIA positions the 4B size for single-GPU and unified-memory systems where a fixed large model will not fit.
Limits: all benchmark figures are NVIDIA's own reported results from the model card, not independent evaluations, and a loop count that maximizes benchmark accuracy costs proportionally more compute per token. The OpenMDW license terms govern commercial use and should be read before deployment. Dory is a base research release in the Nemotron Labs line; real-world coding performance on private codebases is not established by the published scores.