Dory 1.0

NVIDIA's 4B looped model that reruns one shared block 1 to 4 times, so you set its depth when you serve it. OpenMDW-1.1, 256K context.

In short. Dory is a small 4B AI model from NVIDIA that can think harder without getting bigger. For each word it reuses the same group of layers one to four times, and you choose when you start the server whether you want faster answers or better ones. It works in English only.

Summary

NVIDIA published Dory1.0-4B on 16 October 2026 as the first open looped model in its Nemotron Labs line, and a technical report (arXiv 2611.03998) followed in November. It has about 4 billion parameters on a normalized-GPT backbone, with sliding-window and full-attention layers arranged in three groups: an 8-layer input block, a 20-layer recurrent block that is reused, and an 8-layer output block. At serving time you set how many times the recurrent block runs, from one to four, so a single checkpoint covers several compute budgets with no retraining. NVIDIA recommends two loops as the default, one loop to minimize compute and four for agentic work. It was pretrained on 6 trillion tokens, the first 4.5 trillion of them in NVFP4, then extended to a 256K context, and post-trained with supervised fine-tuning and reinforcement learning with verifiable rewards.

NVIDIA's own results compare it with Qwen3.5-4B, Qwen3.5-9B, Gemma 4 26B A4B and Qwen3.6-35B-A3B. On SWE-bench Verified Dory scores 50.3, 59.5 and 61.8 at one, two and four loops, against 54.7 for Qwen3.5-4B and 61 for Qwen3.5-9B. On Terminal-Bench 2.1 it scores 12.5, 18.1 and 22.2, against 15 and 24 for those two. It trails on knowledge tests such as GPQA Diamond, where it scores 69.2 to 72.4 against 77.7 to 85.4. The report says the average over 12 benchmarks rises from 35.2 at one loop to 39.7 at two and 40.6 at four, and that four loops generate 1.3 to 2.5 times fewer tokens than one loop on agentic tests. NVIDIA measured these itself, and the Qwen3.6 SWE-bench score is a reported figure.

The weights use the OpenMDW-1.1 license, and the card says the model is ready for commercial or non-commercial use. The BF16 checkpoint is about 8.0 GB, and NVIDIA also publishes MXFP8 (about 4.8 GB), NVFP4 (about 3.2 GB) and a Q5_K_M GGUF (about 3.7 GB) for llama.cpp. vLLM support was merged on 25 October 2026, while the llama.cpp pull request was still open on 11 November 2026, so the GGUF route needs the pinned build in NVIDIA's instructions. Reasoning is always on in vLLM, and the recommended sampling is temperature 1.0 and top-p 0.95. By default each loop keeps its own KV cache, but a last-loop mode keeps memory at the one-loop size, and NVIDIA says accuracy stays close.

Good for

  • Reasoning, coding and agent tasks on a small memory budget, at about 8 GB in BF16 or 3.2 GB in NVFP4
  • Serving with a speed-versus-quality dial, from one loop for cheap calls to four for harder agent tasks, from the same checkpoint
  • Long inputs of up to 256K tokens, with a last-loop cache mode that keeps KV-cache memory at the one-loop size
  • Researchers studying looped, latent-recurrent models

Watch out for

  • English only
  • More loops mean more compute per token, and the report says fewer generated tokens only partly offset this, so measure latency and cost on your own workload
  • It trails larger rivals on knowledge tests such as GPQA Diamond and Humanity's Last Exam
  • The llama.cpp route needs a pinned build because that pull request was still open on 11 November 2026
  • Benchmark numbers are NVIDIA's own, and its Artificial Analysis scores are internal estimates

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Dory1.0-4B (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Dory1.0-4B (BF16) BF16 4B 9.2 GB 4.6 GB 2.76 GB Laptop or phone
Dory1.0-4B (MXFP8) FP8 4B 9.2 GB 4.6 GB 2.76 GB Laptop or phone
Dory1.0-4B (NVFP4) NVFP4 4B 9.2 GB 4.6 GB 3.52 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.

More about OpenMDW-1.1

See also

Frequently asked questions

What is Dory 1.0?
Dory is a small 4B AI model from NVIDIA that can think harder without getting bigger. For each word it reuses the same group of layers one to four times, and you choose when you start the server whether you want faster answers or better ones. It works in English only.
Can I use Dory 1.0 commercially?
Yes. Dory 1.0 is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Dory 1.0 open source?
Dory 1.0 is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
Can I run Dory 1.0 on my own computer?
Yes. The smallest variant, Dory1.0-4B (BF16), needs about 2.76 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Dory 1.0 released?
Dory 1.0 was released by NVIDIA on October 16, 2026.
How long a context window does Dory 1.0 have?
Dory 1.0 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3.5 Alibaba Qwen 873M – 403B 17B active 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
Nemotron 3 Nano 4B NVIDIA 3.97B 256K NVIDIA Nemotron (Permissive) 2.74 GB Laptop or phone
Nemotron 3.5 Lightning NVIDIA 31.6B 3B active 1M OpenMDW-1.1 (Permissive) 22 GB Gaming GPU