NVIDIA Nemotron (3, 3.5)

Weights for the whole current family are downloadable from Hugging Face, including 3.5 Lightning and 3.5 Bolt in BF16, FP8, and NVFP4 builds.

Permissive in substance but custom in form: the NVIDIA Nemotron Open Model License grants a perpetual, royalty-free, irrevocable right to modify, sublicense, and distribute, and Nemotron 3.5 moves to the Linux Foundation's OpenMDW-1.1, which is permissive but not yet OSI-approved; terms vary per asset.

Deploys on any NVIDIA hardware with open serving stacks including vLLM, SGLang, Ollama, and llama.cpp, from a single-GPU 24GB profile up to the 550B Ultra checkpoint.

Pretraining and post-training datasets (Nemotron-Pretraining-Code-v1, Nemotron-Post-Training-Dataset-v2, Personas, MIND, CrossThink), per-checkpoint training recipes, data cards, and technical reports are published, but coverage is uneven across checkpoints and the exact pretraining mixture for every release is not published.

Stronger than ordinary open-weight releases because recipes and datasets travel with the weights, but reproducing a checkpoint end to end still depends on artifacts that are not published for every model in the family.

NVIDIA publishes Nemotron's datasets and recipes in public repositories that take some outside pull requests, and Nemotron 3.5's OpenMDW-1.1 licence is stewarded by the Linux Foundation, but development direction and releases remain NVIDIA's alone.