Salesforce Unveils In-House Reasoning Model 'Koa' with Nvidia Partnership
Salesforce has developed its first in-house reasoning model, dubbed Koa. The model was created in partnership with Nvidia and is designed to handle mundane but mission-critical tasks such as updating sales records, routing support tickets, and scheduling follow-ups.
Koa is built on Nvidia's open-weight Nemotron 3 Super, which Salesforce post-trained using a proprietary synthetic dataset derived from nearly three decades of CRM experience. The training scenarios span more than 14 industries and cover processes, workflows, and operational policies.
The company claims Koa outperforms leading models while producing three times fewer errors on its internal CRM benchmark. Customer pilots are underway with organizations including 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine, and Xero, with general availability expected in Winter 2026 for U.S. regions.