Nemotron Labs Dory 1.0 4B

Not yet independently verified. Every detail here comes from NVIDIA's own model card, which dates the release to 16 October 2026; no independent report was found yet. The benchmark table on the card is NVIDIA's own measurement, so no score is stored. We will update this when it can be confirmed, and remove this note.

Nemotron Labs Dory 1.0 4B is a small English language model that NVIDIA released on 16 October 2026 as part of its Nemotron Labs line. Its distinguishing idea is looping: instead of a fixed stack of layers, a shared block of 20 layers sits between 8 input and 8 output layers and can be applied several times for each token. NVIDIA describes the model as ready for commercial and non-commercial use, and publishes the weights on Hugging Face under the OpenMDW License Agreement, version 1.1.

What it does

Because the loop count is set at inference time, one checkpoint serves different compute budgets: more loops spend more computation per token for better answers, fewer loops answer faster, and there is no retraining or second model to deploy. NVIDIA says the extra work happens in latent space rather than in longer outputs, so the model needs fewer generated tokens for a task. An option called last-loop KV caching keeps one cache per recurrent attention layer, which NVIDIA says removes the memory overhead that looping would otherwise add. The architecture is a normalized GPT with interleaved sliding-window and full-attention layers, pre-trained natively in NVIDIA's 4-bit NVFP4 format before continuing in BF16, with a context of up to 256,000 tokens.

Who it is for

NVIDIA positions Dory for general-purpose reasoning on a small footprint, which suits teams running models on their own hardware or at the edge who want to trade latency for accuracy per request. The model ships in BF16, MXFP8, NVFP4 and a Q5_K_M GGUF build, and loads with Hugging Face Transformers. The card reports NVIDIA's own results at one, two and four loops against small open models from other vendors on coding, reasoning, instruction-following and long-context tests; those are the vendor's figures under its own harness, and no score is stored here.

Key facts

Maker: NVIDIA Released: 16 October 2026 Size: 4B parameters Context: up to 256K tokens Language: English Formats: BF16, MXFP8, NVFP4, Q5_K_M (GGUF) Licence: OpenMDW License Agreement 1.1