NVIDIA releases Nemotron Labs Dory1.0-4B on Hugging Face, a looped 4B model whose single checkpoint trades inference loops for accuracy (SWE-bench Verified 50.3→61.8 at 1→4 loops, self-reported; 256K context, OpenMDW-1.1)

Why it matters

【Key points】

NVIDIA published Dory1.0-4B, a 4B-parameter looped (latent-recurrent) language model in its Nemotron Labs line, on its official Hugging Face account. The layer stack is split into input, recurrent and output blocks, and the shared recurrent block can run 1 to 4 times: one checkpoint lets you spend more inference compute for more accuracy, with no retraining or model swap. The model card lists an Oct 16 release date, up to 256K context, English only, an nGPT backbone pre-trained natively in NVIDIA's 4-bit NVFP4 format before continuing in BF16, BF16/MXFP8/NVFP4/GGUF (Q5_K_M) formats, and the OpenMDW-1.1 license, which allows commercial use.

On NVIDIA's self-reported numbers, SWE-bench Verified goes 50.3 / 59.5 / 61.8 at 1/2/4 loops (Qwen3.5-9B: 61 in the same table), Terminal-Bench 2.1 rises from 12.5 to 22.2, and GDPval-AA-V2 Elo from 815 to 967. The card recommends 2 loops by default, since most of the gain comes from the second loop. vLLM and llama.cpp support still requires building from unmerged PRs. NVIDIA's Oleksii Kuchaiev endorsed the model on X on Oct 18.

【Verifiable】

Hugging Face model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16

GGUF: https://huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-GGUF

【Our read】

1. The point is the architecture: looping turns inference compute into a dial, and at 4 loops this 4B model's coding scores land near 9B-class models, which matters for on-device and low-VRAM deployment.

2. All numbers are NVIDIA's own and not yet independently reproduced; real latency and cost at higher loop counts need testing per workload.

3. NVIDIA has not announced it on its blog, newsroom or official X account; so far there is only the Hugging Face release and staff endorsement, and runtime support is still at the PR stage.