Local Models: Trust, Control, Optimization — Carter Abdallah, NVIDIA
Open models let builders retain inference traces, customize the training stack, and reduce dependence on one provider, while closed frontier models remain useful for many workloads.
The panel frames open models as ownership of weights, training controls, data provenance, and generated traces. It highlights licenses such as **OpenMDW** that explicitly permit using model outputs to train another model.
For agent systems, preserve traces from open-model runs and treat them as potential fine-tuning data or signals for verifier-driven training. Consider a mixed stack: owned models where control and continuity matter, closed frontier APIs where convenience or capability wins.
The panel frames open models as ownership of weights, training controls, data provenance, and generated traces. It highlights licenses such as **OpenMDW** that explicitly permit using model outputs to train another model. For agent systems, preserve traces from open-model runs and treat them as potential fine-tuning data or signals for verifier-driven training. Consider a mixed stack: owned models where control and continuity matter, closed frontier APIs where convenience or capability wins. Open does not make operation free or automatically better. The speakers expect open and closed systems to coexist, and the claimed **4-billion-parameter phone model** comparison was a panel assertion rather than a presented benchmark.
This expands the open-model decision beyond serving cost to ownership of weights, provenance, training controls, continuity, and reusable inference traces. Those traces may become training assets where licenses permit, making data governance part of model strategy. It supports coexistence with closed APIs, while offering no benchmark that ownership, local operation, or the cited phone-scale model improves task outcomes.