Nvidia releases two open AI models for agents and announces a third

Nvidia has released two new open AI models in the Nemotron 3.5 family. A third, the smallest, is only planned for now. The models are aimed at AI agents, meaning programs that solve tasks on their own in multiple steps, such as finding information, using tools and delivering a result.

The big model: Super VL

Nemotron 3.5 Super VL has 120 billion parameters, which are roughly the model's adjustable knobs. According to Let's Data Science, it was released around October 15, 2026, and is designed for specialized agents that run continuously.

The model is built as a mixture-of-experts. That means only part of it is used at a time: here, 12 billion parameters are activated per token (a token is a small piece of text). This saves computing power. According to Nvidia, it can reason, follow instructions and use tools, and it understands text, images, documents and video. Nvidia says it was trained on roughly 26 trillion tokens.

Nvidia claims up to five times higher throughput on agent tasks than other open models of the same size and four times faster task completion than comparable models. Nvidia also reports 63.44 percent on Hermes Bench and 80.8 percent on the chart test CharXiv. These are the manufacturer's own figures. According to Let's Data Science, the model scores 26.5 on the independent Artificial Analysis Intelligence Index.

The weights, meaning the trained model itself, are available on Hugging Face and ModelScope. It can also be used via OpenRouter and as an Nvidia NIM microservice. Nvidia names Scale AI, Ramp and AI21 among its partners.

The local model: Bolt 74B-A9B

Nemotron 3.5 Bolt 74B-A9B is built to run locally. According to Let's Data Science, it uses 2-bit quantization, a form of compression, so it can fit in 24 GB of memory. It is meant for research, coding and multistep tasks on Nvidia RTX and DGX Spark and is available now. It scores 21.4 on the Artificial Analysis index, and Nvidia says it ranks first among open models in its class on PinchBench.

The third: Bolt E4B

Nemotron 3.5 Bolt E4B takes up 4 GB and is intended for fast conversational agents, voice assistants and embedded interfaces. According to Let's Data Science, it arrives on October 23.

According to IT Brief, local operation can mean less dependence on the cloud for sensitive tasks for businesses. The license terms are not stated in the sources.