Nvidia’s Dory 4B reruns its own layers to think harder, and its weights are open

Nvidia has published NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 on Hugging Face, a 4B-parameter model that reapplies a shared block of layers up to four times, so users can trade speed for accuracy without retraining.

Three layer groups with a loop count set per request

Dory is what Nvidia calls a looped, or latent-recurrent, language model. Instead of stacking many distinct layers, its design splits the network into three groups: 8 input layers, 20 recurrent layers that are shared and run again on each pass, and 8 output layers. The input is fed back in at every loop. According to the model card, the number of loops can be set to 1, 2, 3 or 4 at inference time. Attention alternates between sliding-window and full-attention layers.

The model supports context of up to 256K tokens, which Nvidia says it validated in vLLM. It handles text only, in English, and its training data also covered 43 programming languages. Nvidia lists tool calling as supported.

On training, Nvidia says the model saw 6T tokens in pre-training, using NVFP4 precision for roughly the first 4.5T tokens and BF16 for the remainder. A further 300B tokens extended the context window to 256K. Post-training combined supervised fine-tuning with GRPO reinforcement learning across math, code, science, instruction following, tool use, multi-turn conversation and structured output. Some synthetic data came from teacher models such as GPT-OSS-120B, the card says.

The loop count is the main lever. In benchmarks that Nvidia says it measured itself, scores rise as loops are added:

• SWE-bench Verified: 50.3 with one loop, 59.5 with two, 61.8 with four.
• Terminal-Bench 2.1: 12.5, 18.1 and 22.2 respectively.
• GDPval-AA-V2 (Elo): 815, 946 and 967.
• IFBench: 74.29, 77.94 and 79.

Gains are not uniform. On GPQA Diamond the two-loop score (72.41) is slightly above the four-loop one (72.16), and on AA-Omniscience accuracy the score stays near 10 whatever the loop count. Nvidia’s comparison table also lists Qwen3.5 and Gemma 4 models, and the larger ones lead on several knowledge-heavy tests such as HLE and GPQA Diamond. Nvidia recommends two loops as the default, with temperature 1.0 and top_p 0.95. The page also says looping reduces the number of tokens the model needs to generate to solve tasks.

Nvidia is also offering quantized versions: MXFP8, NVFP4 and a GGUF Q5_K_M build for llama.cpp. On a DGX Spark at batch size 1, it reports speedups over BF16 of 1.69x to 1.84x for MXFP8 and 2.17x to 2.44x for NVFP4, depending on input and output length.

Availability and license

The weights are released under the OpenMDW License Agreement v1.1, which the page says allows commercial and non-commercial use. The card lists Hopper (H100) and Blackwell (B200, B300, DGX Spark) GPUs on Linux. Serving currently needs a vLLM pull-request build or a Dory-enabled llama.cpp build, since neither has shipped the support in a release yet. The model card lists a release date of October 16, 2026.

A small model that trades loops for size

Most of this week’s open-weight news has been about size. Meta’s Muse Spark Open has 816B parameters, and Reflection’s Beam has 501B under Apache 2.0, according to our earlier coverage. Dory goes the other way: a 4B model with only 36 distinct layers, where extra accuracy comes from running compute again on the same weights rather than from storing more of them.

That has a practical consequence for anyone deploying on a single GPU or a DGX Spark. Memory footprint is fixed by the 4B parameter count, while the loop setting gives a dial between latency and quality at request time. A coding assistant could run one loop for quick completions and four for harder SWE-bench style tasks. The numbers above suggest the dial matters most on agentic and coding work, and much less on factual recall, which fits the idea that repeating computation helps reasoning more than it adds stored knowledge.

It also fits Nvidia’s wider open-model push. The company’s $12.9 billion bet on Hugging Face, covered in September, makes the platform a natural place to release research models under the Nemotron Labs name. The caveats are Nvidia’s own: the scores come from its evaluation harness, serving needs unreleased framework builds, and the card says the training data may underrepresent some demographic groups and recommends bias audits.