Nvidia Releases Open-Weight Bolt Model for Local AI Agents

Nvidia has released an open-weight AI model that runs in 24GB of memory on compatible PCs and systems, allowing users to operate agents locally on their own hardware.

Nemotron 3.5 Bolt 74B-A9B 24GB became available Oct. 15. The 74-billion-parameter model is designed for agents that handle multistep reasoning, use software tools and browse the web.

Supported hardware includes AI PCs running Nvidia RTX Spark and DGX Spark workstations. Running the model locally lets users carry out those workloads without sending them to the cloud.

Open-weight models give developers access to the parameters that shape a model’s output, so they can adapt it and select where to deploy it. Bolt brings that option to local systems. The release follows Nvidia’s expansion of RTX Spark software for running AI models on local machines.

Nvidia also made Nemotron 3.5 Super VL available Oct. 15. The 120-billion-parameter model is built for enterprise agents working across reasoning, tool use and visual documents. Companies can train it further on their own data and deploy it in their data centers.

Ramp is among the companies using Super VL. CrowdStrike is among the companies using Super VL.

The releases give organizations two ways to run customized agents on their own infrastructure. Super VL is aimed at larger enterprise deployments, while Bolt is designed to fit within the memory available on supported local systems.

Nvidia plans to release Nemotron 3.5 Bolt E4B later in October. The smaller version is intended for conversational agents and other local applications.