Salesforce has unveiled Koa, a new AI reasoning model, at its Dreamforce tech conference. Built on Nvidia's open-weight Nemotron model, Koa has been post-trained by the two companies to specialise in sales, marketing, and customer support tasks.
Koa is presented as an open-weight alternative to proprietary frontier models. Salesforce states that Koa has not ingested any actual customer data, aiming to prevent data leaks, and is designed to reduce AI spending by using fewer tokens for tasks.
Jayesh Govindarajan, EVP of Salesforce AI, noted that while Salesforce has developed task-specific language models, they previously relied on frontier model providers for reasoning capabilities. Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, highlighted Nemotron's architecture for token-efficient inference.
Koa will be offered within Salesforce's Agentforce platform, where customers create agents for routine tasks. Salesforce also recently announced a partnership with Anthropic, called ClaudeForce, which allows companies to use Claude as an AI interface while their data remains within Salesforce's systems.