Nvidia Unveils Open-Weight Nemotron Models for Enterprise Agents and Local AI PCs

Nvidia has added two new families to its open-weight Nemotron lineup, targeting both data-center agent deployments and the emerging class of AI PCs. The flagship Nemotron 3.5 Super VL packs 120 billion parameters and is engineered for always-on enterprise agents that handle document intelligence, financial accounting, scientific experimentation, insurance claims processing, and security resilience tasks. Despite being smaller than the Nemotron 3 Ultra, Nvidia says post-training on agentic benchmarks yields comparable instruction-following and tool-use accuracy while completing tasks faster than several proprietary frontier models.

On the edge side, the company introduced two Bolt variants built to run locally on PCs powered by the new RTX Spark chip and on DGX Spark workstations — hardware now shipping through partners including Microsoft and Dell. The 74-billion-parameter Bolt 74B-A9B 24GB balances latency and reasoning for multi-step tool calling and web browsing, while the smaller Bolt E4B targets conversational assistants and will arrive later.

Why it matters for GPU / AI infrastructure

Open-weight models that organizations can post-train on proprietary data and deploy on owned infrastructure reduce dependency on closed APIs and give IT teams full control over data residency and compliance. For AiGpu customers, the Super VL’s optimization for Nvidia’s hardware stack means higher throughput per GPU dollar on existing clusters, while the Bolt series validates the economics of running agentic workloads on workstation-class GPUs instead of scaling to cloud instances. As frontier labs diversify across Trainium, Cerebras, and custom silicon, Nvidia’s open-model strategy strengthens the case for standardizing on its accelerated platform from cloud to edge.