Salesforce's Koa brings reasoning in-house on Nvidia's Nemotron

Salesforce introduced Koa at its Dreamforce conference as its first reasoning model, built on Nvidia's open-weight Nemotron. The company says it had long wanted to train its own enterprise-grade frontier model but lacked a suitable pre-trained base. Jayesh Govindarajan, EVP of Salesforce AI, pointed to Nemotron as the first sovereign American pre-trained model that was state of the art and had clear data provenance, contrasting it with Alibaba's Qwen, whose training data he said Salesforce cannot know.

To build Koa, Salesforce and Nvidia avoided actual customer data, instead crafting synthetic data that simulated a customer service environment, from irate callers to a sales professional closing a deal. Koa is meant to be better and cheaper in tokens burned at the work tasks Salesforce customers want an agent to do, compared with sending those same tasks to Claude or ChatGPT. Nvidia's Kari Ann Briski described Nemotron's inference architecture as token efficient, framing sovereign AI, time to first token and efficient reasoning as the combination that matters.

The move sits alongside Salesforce's continued reliance on the frontier labs: the company also announced ClaudeForce, a partnership with Anthropic that lets companies use Claude as their AI interface while data stays in Salesforce's system of records. The test for Koa is likely to be whether enterprises accept a task-specific, open-weight reasoning model in place of the frontier models they have routed work to until now, and whether the token savings and data control outweigh any capability gap.