Salesforce launches Koa, a reasoning model built on Nvidia's open Nemotron

Salesforce introduced Koa, its first reasoning model, this week at its Dreamforce conference. The model is built on Nvidia’s Nemotron, an open-weight model whose code and parameters are publicly available, and the two companies trained it together to handle sales, marketing and customer-support work inside Salesforce’s Agentforce platform, where customers build automated agents for tasks like answering support tickets.

Before Koa, Agentforce routed any task requiring multi-step reasoning to outside frontier models such as Anthropic’s Claude or OpenAI’s ChatGPT, through what Salesforce calls an AI gateway, a system that decides which model handles which request. Jayesh Govindarajan, Salesforce’s EVP of Salesforce AI, told TechCrunch the company had built smaller task-specific models before but had always relied on frontier labs for reasoning until Nemotron gave it a base model to build on. He said Nemotron was the first option that was a “sovereign American pre-trained model” that was also state of the art and had clear data provenance, contrasting it with Alibaba’s Qwen, whose training data he said is unknown.

To adapt Nemotron for CRM work, Salesforce and Nvidia did not use any real customer data. Govindarajan said the companies instead built synthetic training data by simulating a customer-service environment with fictional personas, including irate callers and sales professionals trying to close deals.

Salesforce says Koa is both better suited to these work tasks and cheaper to run than sending the same requests to a frontier model, because it uses fewer tokens, the units of text a model processes, to complete them. Kari Ann Briski, Nvidia’s vice president of generative AI software for enterprise, told TechCrunch that Nemotron’s inference architecture was built for token efficiency, calling sovereign AI, fast response times and efficient reasoning “the trifecta of things that you need to have” for controlling AI costs.

Still working with Anthropic

The launch does not mean Salesforce is stepping away from the frontier labs entirely. The company also announced a partnership with Anthropic called Claudeforce, which lets customers use Claude as an interface for Agentforce while their data stays inside Salesforce’s own systems and security infrastructure.

Koa will now sit alongside Claude and ChatGPT as an option inside Agentforce’s AI gateway, giving customers a choice between an in-house model tuned for CRM tasks and general-purpose frontier models for other work. Salesforce has not disclosed pricing for Koa or said when it will be broadly available to all Agentforce customers, details worth watching as the company moves the model from its Dreamforce announcement into production use.