Salesforce and Nvidia Launch Koa AI Model

Salesforce used its Dreamforce conference in San Francisco this week to unveil Koa, its first reasoning model, built on Nvidia's open-weight Nemotron architecture. The two companies post-trained it specifically for sales, marketing, and customer-support work rather than general problem-solving.

That distinction matters more than it sounds. OpenAI, Anthropic, and Google want enterprises feeding files, code, and feedback directly into their proprietary systems, often at steep per-token cost. Koa runs the other way: it's open-weight, hasn't touched actual customer data, and uses fewer tokens to complete the same tasks. Salesforce says it can also route automatically through an AI "gateway" depending on the job, while staying inside whatever data and security rules a given customer already has in place.

Koa will sit inside Agentforce, Salesforce's platform for building agents that handle things like answering support tickets or booking appointments, as an alternative to the other models already available there. Jayesh Govindarajan, the company's EVP, frames it as part of a broader shift away from one-size-fits-all frontier models toward smaller, task-specific ones built for the grind of enterprise work rather than benchmark math problems.