NVIDIA's Nemotron 3.5 puts open agent models on a size spectrum.

NVIDIA introduced three additions to its Nemotron family on October 15: the 120-billion-parameter Super VL model for larger agentic tasks, Bolt 74B-A9B aimed at local systems with a 24 GB memory footprint, and a still-smaller Bolt E4B announced for release on October 23. The first two were available at announcement; the smallest was not yet released.

The company emphasizes specialized workflows, visual understanding, tool use and local execution. Its 'five times faster' comparison for Super VL is based on the stated benchmark setup and model class. That result should not be generalized to every workload.

Agent systems do not need the same model for every action. Routine decisions, document retrieval, visual understanding and complex reasoning have different latency, privacy and cost requirements. A model family covering several hardware tiers could encourage builders to design systems around the job rather than insist on one huge model.

Open weights can also give organizations greater control over deployment and customization. But that control brings operational responsibilities: model evaluation, security updates, monitoring, hardware provisioning and clear policies for tool use.

The strategic signal is modular AI. The most economical agent may route work between smaller local models and stronger remote or server-hosted ones, with a human review point for consequential actions.

Cortist would test a repeatable production task set using three arrangements: a small local model, a larger local or private model, and a hosted model. Compare success per dollar and intervention hours, not only leaderboards.