NVIDIA-Nemotron-Labs-Dory1.0-4B
NVIDIA released Nemotron-Labs Dory1.0-4B, a 4B text model on Hugging Face in BF16, MXFP8, NVFP4 and GGUF versions.
NVIDIA published a small 4B model under its Nemotron-Labs line, with quantized checkpoints for lower memory use. On DGX Spark, NVFP4 ran 2.17x to 2.44x faster than BF16 and MXFP8 ran 1.69x to 1.84x faster. The model card lists training data from the Nemotron pre-training and post-training v3 datasets, and the chat template has a thinking mode.
- Date
- Friday 16 October 2026
- Lab
- NVIDIA
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| DGX Spark speedup vs BF16, NVFP4 (W4A16) | 2.17x to 2.44x batch size 1, input 2560 and output 300 tokens, and input 128 and output 1024 tokens | company |
| DGX Spark speedup vs BF16, MXFP8 (W8A16) | 1.69x to 1.84x same settings | company |
Released under the openmdw-1.1 license. Benchmark scores and the architecture were not visible in the page text I read.
Sources
- huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16
- huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-NVFP4
- huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-MXFP8
This record was partly confirmed: some claims could not be checked on 18 October 2026. How we check