NVIDIA has added a new large language model called Dory1.0‑4B to Hugging Face. It has 4 billion parameters and can handle contexts up to 256 K tokens. The model uses a looped (latent‑recurrent) architecture that lets users trade compute for accuracy without retraining. It is offered in several formats, including BF16 and 4‑bit NVFP4, and can be used for commercial or non‑commercial projects under the OpenMDW‑1.1 license.
Why it matters
Developers can run a single 4B‑parameter model on varying hardware while still supporting very long prompts, reducing the need for multiple specialized models.
