NVIDIA releases Nemotron 3.5 Super VL, an open-weight model that runs 5x faster than competitors

NVIDIA has just announced 3 new open weight AI models under the Nemotron 3.5 umbrella. Why does this matter? It allows developers to leverage new tools to build their own autonomous agents right now, without having to wait for access to proprietary APIs.

New open-weight models enable local AI agent deployment

This includes a large 120B parameter model for advanced coordination. A mid-sized 74B parameter model for on-prem deployment. And a tiny 4B parameter model for the edge. So, now people can choose from a range of sizes when they're trying to deploy AI agents on their own hardware and need to work with certain constraints.

Central to this is the Nemotron 3.5 Super VL. A model NVIDIA created for advanced coordination of always-on multimodal agents. This model performs at the same level as NVIDIA's biggest reasoning model, but with just about 1/5 of the parameters. This model is claimed to have best-in-class accuracy for multimodal agentic tasks and runs 5x faster than the closest competitor in the open weight category.

NVIDIA also announced the Nemotron 3.5 Bolt 74B-A9B 24GB for local deployment. A 74B parameter model that only uses 9B parameters at any given time, by applying 2-bit quantization and squeezing them into a 24GB of memory. This makes it optimal for local agentic workloads that can be executed on consumer grade GPUs with low amounts of VRAM. So it strikes a good balance between computational depth and memory for the purpose of making advanced agentic reasoning possible in smaller scale setups.

Lastly, NVIDIA announced the Nemotron 3.5 Bolt E4B 4GB for high-speed physical and conversational agents. A model with only 4B parameters, created specifically by NVIDIA for fast and low resource use cases on the edge. These 3 models can now be found on Hugging Face.

They're compatible with things like vLLM, Ollama, LM Studio and more. So you know that you can easily integrate these models into what you already do without needing to change too much.

NVIDIA just announced these 3 new open weight models. They give you options for different agentic workloads. So you've got high performance cloud coordination all the way down to lightweight deployment on the edge. Now you can access all these models without needing to go through any hoops, directly from regular open source repos and popular inference tools.