Until Koa, the multistep reasoning behind Salesforce's agents ran on Claude or ChatGPT.
Salesforce and Nvidia announced Koa at Dreamforce. It is available to select pilot customers through Agentforce, Salesforce's platform for building AI agents, where a gateway picks the model for each request. General availability in U.S. regions is expected in winter 2026.
The two companies post-trained it from Nvidia's open-weight Nemotron 3 Super for selling, marketing and customer-support work; Koa's weights stay with Salesforce, which runs the model inside its own infrastructure.
Jayesh Govindarajan, Salesforce's executive vice president for AI, told TechCrunch that the company had built many small models for specific Agentforce jobs. He cited training-data provenance as a reason for choosing Nemotron, and said Salesforce did not know what data Alibaba's Qwen was trained on.
No customer data went into the post-training. The companies built synthetic stand-ins instead, simulating interactions from furious callers to a salesperson pushing to close a deal.
Salesforce's own tests, reported in the paper published with the model, put Koa's overall CRM Bench score above GPT-4.1's and below those of Claude Opus 4.8 and GPT-5.5.
In August Salesforce announced Claudeforce, a partnership with Anthropic under which companies can use Claude as their AI front end while their data stays in Salesforce's system of record.