NVIDIA releases three new open Nemotron 3.5 models

By Splitfeed EditorialLast updated Oct 15, 2026, 15:10 UTC

On October 15, 2026, NVIDIA released three models under its Nemotron 3.5 family. The largest is Nemotron 3.5 Super VL, an open 120-billion-parameter model positioned for always-on specialized agents, emphasizing task completion speed and accuracy. The two smaller releases are edge-focused: Nemotron 3.5 Bolt 74B-A9B 24GB, which targets local agentic workloads by balancing intelligence and speed per gigabyte, and Nemotron 3.5 Bolt E4B 4GB, designed for deployment in environments where memory is constrained to as little as four gigabytes.

The release is notable because it attempts to close a gap that has separated high-capability AI models from consumer and edge hardware. By spanning from a 120-billion-parameter data center model down to a four-gigabyte edge variant under a single product family, NVIDIA is signaling that agentic AI workloads, those in which a model autonomously completes multi-step tasks, need not be confined to cloud infrastructure. What remains genuinely in dispute is whether open models at these sizes and memory footprints can deliver the reliability that real-world agent deployment requires, and whether NVIDIA's efficiency claims hold up across varied hardware configurations beyond its own ecosystem.

No reactions from the Pro-AI, Anti-AI, or Middle Ground camps have been published yet. Typically, the Pro-AI camp would highlight the democratizing effect of open, edge-deployable models and frame the release as evidence that capable AI is becoming widely accessible. The Anti-AI camp would likely raise concerns about the proliferation of powerful agentic models running locally, where oversight and safety controls are harder to enforce. The Middle Ground camp would probably welcome the openness and efficiency gains while pressing for independent benchmarks to verify the performance and safety characteristics NVIDIA is claiming.

The argument will sharpen once third-party researchers publish benchmark comparisons against rival open models such as Meta's Llama family and Mistral's edge offerings, and once developer communities report real deployment experiences on constrained hardware. Independent evaluations of the Super VL model's accuracy claims on agentic tasks will be the most concrete signal to watch.