NVIDIA Releases Nemotron-Labs Dory 4B BF16 Model on Hugging Face
NVIDIA has released the Nemotron-Labs Dory1.0-4B-BF16 model on Hugging Face, introducing a small language model with unique architectural features. The model utilizes a looped structure allowing users to adjust compute budgets dynamically at inference time. This release gives developers and researchers access to a novel, highly efficient 4-billion-parameter model capable of competitive performance on complex coding benchmarks. It offers flexible compute scaling, which represents a notable design shift for open-source language models. The model reuses its middle layers multiple times, with inference loop counts that can be configured by the user. NVIDIA reports strong benchmark results, such as 61.8 on SWE-bench Verified when using four loops.
## BACKGROUND
BF16 (Bfloat16) is a 16-bit floating-point data type commonly used in machine learning to speed up training and inference while preserving model accuracy. Hugging Face is a widely used platform where the machine learning community shares datasets, pre-trained models, and application demos.