Salesforce and NVIDIA Unveil Koa, Their First CRM Reasoning Model Built on Nemotron

My take: Salesforce and NVIDIA unveiled Koa at Dreamforce, the first reasoning model designed specifically for CRM data. It was built by post-training NVIDIA's Nemotron-3-Super (120 billion parameters) on synthetic data drawn from 27 years of Salesforce CRM deployments across more than 14 industries. On Salesforce's internal benchmark, Koa makes 3 times fewer errors than general-purpose models on CRM tasks like updating an opportunity, routing a case, or scheduling a follow-up.

It is worth noting that the benchmark was run by Salesforce on tasks from its own platform, which gives it an obvious bias in favor of the model. The figures should be read critically until independent evaluations are available. Even so, the direction matters: domain-specialized models are beginning to outperform general-purpose models on the tasks they were built for, and that changes how enterprise agent systems get built.

For teams that work in Salesforce, Koa is not yet generally available; it is currently in pilot with selected customers, with general availability expected this winter. The question for operations and technology leaders: are you evaluating whether your current AI tools are the right fit for your specific business tasks, or are you still using a single model for everything?