Salesforce and Nvidia Unveil Koa: A New AI Reasoning Model Challenging Frontier AI

Salesforce and Nvidia have collaborated on a new AI reasoning model called Koa, which challenges frontier AI models by providing an open-weight alternative that is trained to perform specific work tasks rather than generic problems. The model has been designed to address the concerns of companies who pay twice for using AI models - once monetarily and again by giving away company information that may be used for training.

Koa is built on Nvidia's Nemotron chips and is tuned for multi-step sales, service, and marketing workflows. According to Marc Benioff, Chair and CEO of Salesforce, the model has been trained with synthetic scenarios that reflect real-world enterprise workflows across 14 industries, including manufacturing, financial services, healthcare, and travel.

The model can reason through long-running or multi-step tasks without relying on external AI models, making it a more secure and controlled option for companies. Nvidia's Jensen Huang noted that the open-weight model gives Salesforce the foundation to turn decades of enterprise expertise into specialized AI with Koa, creating a CRM model that can reason and securely take action.

Koa is currently being offered as an alternative to other models in Salesforce's AgentForce platform, which is used for building agents to handle regular tasks such as answering customer service queries and scheduling appointments. The company has announced a partnership with Anthropic to allow companies to use Claude as their AI interface while keeping data on Salesforce's systems.

For now, Koa will be made available only to select pilot customers via Agentforce, but it is expected to be rolled out over the next few months in the United States and thereafter across the rest of Salesforce's customer base.