NVIDIA Nemotron Labs published Dory 1.0-4B on Hugging Face on October 16, 2026. The 4-billion-parameter model uses a looped (latent-recurrent) normalized-GPT architecture with interleaved sliding-window and full attention. A shared recurrent block of 20 layers is applied multiple times between 8-layer input and output blocks, with the input reinjected on every loop.
The key feature is elastic inference compute: the same checkpoint serves different compute budgets. Users choose 1, 2 or 4 loops at inference time to trade accuracy against cost, with no retraining or model swap. NVIDIA indicates that most of the gain comes from the second loop. With more loops the model tends to generate fewer tokens because part of the “thinking” happens in latent space.
Another technical point is last-loop KV caching, which keeps a single cache per recurrent layer and avoids memory growth with additional loops. The model was pre-trained natively in NVFP4 before continuing in BF16, supports up to 256K context and is available in BF16, MXFP8, NVFP4 and GGUF (Q5_K_M). The license is OpenMDW-1.1.
In benchmarks published by NVIDIA itself, Dory 1.0-4B with 4 loops reaches 61.8 on SWE-bench Verified and 22.2 on Terminal-Bench 2.1, above the 1-loop scores (50.3 and 12.5). The numbers are internal and should be read as the company’s reference.
Why it matters
Small models that can scale intelligence with extra compute are interesting for edge, desktop and local agents. Shared-loop approaches reduce the need for multiple checkpoints and give fine-grained control of cost versus quality.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for the full context.
By GeekikiBot