AI News
Nvidia Goes Fully Open: New Nemotron Models Ship With Weights, Datasets and Recipes
Nvidia released a new family of Nemotron 3.5 models this week, and the headline isn't just the models — it's how open they are. The release ships with weights, datasets, AND training recipes, making it one of the most complete open AI releases from a major lab to date. The models are available on Hugging Face.
Releasing the training recipes alongside the weights and data goes a step beyond the typical "open weights" release. It means developers and researchers don't just get a finished model — they get the instructions for how it was made, which makes the work reproducible, auditable, and easier to build on.
Three models, three use cases
The family spans three models aimed at different workloads. Nemotron 3.5 Super VL, at 120 billion parameters, is designed for "always on" agents — persistent assistants that run continuously rather than answering one-off queries. Nemotron 3.5 Bolt 74B-A9B 24GB is the balanced option, trading off size and capability for general-purpose use. And Nemotron 3.5 Bolt E4B 4GB is built for conversational agents, small enough to run in constrained environments.
The range shows Nvidia targeting the full spectrum of agent deployment: from heavyweight always-on systems down to lightweight conversational models that could run on-device or on modest hardware.
An on-device and open-source strategy
The release reflects a strategy that pairs open source with on-device deployment. By publishing everything needed to reproduce and adapt the models, Nvidia is betting that openness will drive adoption of its models across the ecosystem — including on hardware far beyond its own data-center GPUs.
The timing is notable. With Meta releasing the open-weight Muse Spark the same week, the open AI ecosystem is having a moment: two of the biggest names in AI are simultaneously pushing powerful models into the open, challenging the closed-model dominance of rivals. The full recipe disclosure also lowers the barrier for fine-tuning and domain adaptation, since teams can see exactly how the base model was built before they customize it.