Salesforce and Nvidia launch Koa reasoning model for enterprise AI

At its Dreamforce conference, Salesforce introduced Koa as its first reasoning model, built on Nvidia's open-weight Nemotron and post-trained with Nvidia specifically for sales, marketing, and customer-support tasks. Until now, Salesforce relied on frontier model providers for reasoning, routing prompts to models like Claude or ChatGPT through Agentforce's AI gateway when an agent needed to work through a multi-step task. Koa is meant to handle those same tasks inside Salesforce's own stack.

Jayesh Govindarajan, EVP of Salesforce AI, said the company had always wanted to train its own enterprise-grade frontier model but lacked a suitable pre-trained base. In his words, until Nemotron there was no sovereign American pre-trained model that was available, state of the art, and had clear data provenance. He pointed to uncertainty about what Alibaba's Qwen trains on. Nvidia's Kari Ann Briski framed the combination as sovereign AI, time to first token, and efficient reasoning, for what she called the tokenomics.

The pitch to enterprise buyers rests on several claims: an open-weight option instead of closed frontier models, training on synthetic data rather than actual customer data, fewer tokens for the same work, and automatic routing through an AI gateway. Salesforce isn't abandoning Anthropic or OpenAI; it also announced Claudeforce, a partnership letting companies use Claude as their AI interface while data stays in Salesforce's system of records. Whether Koa's token savings and data-provenance guarantees hold up in practice is likely to decide how readily enterprises move reasoning work off frontier providers.