The newly debuted Koa model leverages an open-weight Nvidia Nemotron-3-Super-120B base to automate reasoning for complex enterprise business processes. Now in pilot for the Agentforce platform, the model reduces errors on CRM actions three-fold compared to leading performers without using customer data for training. It orchestrates a portfolio of domain-specialized models to handle specific tasks like toxicity screening and search reranking.

Technical benchmarks show Koa scoring 69.41 on Tau2Bench, beating GPT-4.1's 54.48, and reaching 0.86 on the CRM Bench. Training utilized Group Relative Policy Optimization and converted Agent Script specifications into simulated multi-turn tasks to increase function-call accuracy from 0.71 to 0.77. This collaboration also brings Nvidia accelerated computing to Missionforce for mission-specific AI agents.