News · 2026-10-17

NVIDIA released Dory, a small model that reuses its layers to spend more compute

NVIDIA released Dory 1.0, a four-billion-parameter language model that can repeat a shared block of layers to spend more computation on each token. The same checkpoint supports different loop counts without retraining. Its release makes adjustable inference depth available as a documented model feature, although NVIDIA's own tests show that additional loops do not erase every gap with larger models.

Key facts

• NVIDIA dates Dory 1.0 to October 16, 2026 and marks the model generally available.
• The architecture has eight input layers, a shared 20-layer recurrent block and eight output layers.
• The card lists a 256,000-token context window and English support.
• The primary source is NVIDIA's Dory model card.

A conventional language model sends its internal representation through a fixed stack of layers. Dory introduces a reusable middle section. Running that section again gives the representation another round of processing while keeping the model's stored parameters unchanged. The model can therefore use more arithmetic without requiring a separate, larger checkpoint for every operating mode.

The closest everyday analogy is rereading a document with the same team rather than hiring a second team and keeping another copy of all its instructions. More passes can expose relationships missed on the first pass. They also take time, and repeating a weak interpretation does not guarantee a good answer. Adjustable depth creates a resource tradeoff; it does not make extra work universally useful.

NVIDIA documents one-, two- and four-loop settings. The published card describes a latent-recurrent normalized architecture combining local-window and full attention. Those terms concern how information is updated and which earlier positions can be consulted. The practical point is simpler: a user can select how often the central computation repeats while loading the same underlying model.

Memory is an important part of that design. Models normally retain previously computed attention information so that every new token does not require rebuilding the whole conversation. Ground Truth's cache lesson explains this storage. In a recurrent architecture, maintaining separate cached information for multiple loops can increase memory use as loop count rises.

Dory offers an optional last-loop caching mode. NVIDIA says this keeps the cache from growing with the number of recurrent passes when enabled. That claim concerns one component of runtime memory, rather than the total footprint. The model still needs its weights, temporary computation space and conversation state. A cache optimization is not a complete minimum-hardware specification.

The release lists several precision and packaging formats, including BF16, MXFP8, NVFP4 and a quantized GGUF option. These labels describe different ways of storing and computing with numerical values. The dossier does not verify an individual checkpoint download size or an explicit minimum graphics-memory requirement, so neither is supplied here. A four-billion-parameter description alone cannot answer those questions.

Licensing is more definite. NVIDIA lists OpenMDW License Agreement 1.1 and says the model is ready for commercial and non-commercial use. The NVIDIA collection listings supply broader publisher context. The Nemotron-Labs label should not be mistaken for an unreleased preview: the card explicitly marks the version as “GA (10/16/2026).” The license and card remain the authoritative documents for the scope of permitted use.

The strongest technical counterargument appears in NVIDIA's own results. More loops improve some measured scores, yet larger comparison models remain ahead on several reasoning, coding and agent tasks. Spending extra computation within a compact parameter budget is useful for studying efficiency; it is not evidence that a small recurrent model has become equivalent to the strongest frontier system.

The measurements are also vendor-run. NVIDIA describes a shared evaluation harness, which helps interpret comparisons, but the dossier establishes no outside replication of this release. Hardware speedups for compressed variants are similarly attached to NVIDIA's tested system and baseline. A prospective user should compare completed-task quality and latency at each loop setting, rather than assume four loops is always the best choice.

One community headline complained that a Nemotron model was too guarded for useful agent work. The local findings identify that account as involving a different, much larger model, and the stronger dossier pass does not verify its post body. It therefore cannot establish that Dory itself refuses a particular workflow. Model-family branding is too broad a unit for attributing specific behavior.

The broader concept is covered in Ground Truth's looped-transformer lesson. Dory makes it concrete as a released checkpoint with configurable recurrent depth. Its significance lies in exposing computation as a selectable serving choice, while leaving the actual quality, latency and memory tradeoff to be measured for each use case.

Primary source, verified: read the paper →

Ground Truth. (2026, October 17). NVIDIA released Dory, a small model that reuses its layers to spend more compute. Ground Truth. https://groundtruth.day/news/nvidia-dory-adjustable-recurrent-depth.html

@misc{groundtruth:nvidia-dory-adjustable-recurrent-depth,
  title  = {NVIDIA released Dory, a small model that reuses its layers to spend more compute},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/nvidia-dory-adjustable-recurrent-depth.html}
}

Topics: model-release · inference · architecture · open-weights