Salesforce and NVIDIA announced Koa on September 15, a CRM reasoning model for Agentforce built by post-training NVIDIA Nemotron 3 Super. Salesforce says its training used synthetic enterprise scenarios spanning more than 14 industries, rather than customer data.
The model targets multi-step CRM actions such as updating opportunities, routing support cases and scheduling follow-ups. Salesforce controls the weights and says training and inference run within its own infrastructure. Its reported reduction in errors comes from a Salesforce benchmark; that result is a vendor evaluation and does not establish performance across every customer's workflows.
Koa is available to selected pilot customers, with general availability expected in Winter 2026 in U.S. regions. That is a planned rollout, not a worldwide launch. For Agentforce customers, the practical question is whether a specialized model improves the accuracy of their own tool calls and business processes. Pilot testing should use representative tasks and failure cases before relying on the headline benchmark.