Salesforce and Nvidia unveil Koa, an open-weight reasoning model for enterprise tasks
Salesforce introduced Koa, its first reasoning model built on Nvidia's open-weight Nemotron base, at Dreamforce. The model is post-trained for sales, marketing, and customer-support tasks, offering enterprises an alternative to closed frontier models with lower token usage and no customer data ingestion. It will be available through Salesforce's Agentforce platform, with routing via an AI gateway.
Koa was developed through a joint effort where Salesforce post-trained Nvidia's open-weight Nemotron base model using synthetic data that simulated realistic customer service scenarios, including interactions with difficult customers and sales professionals working to close deals. No actual customer data was used in the training process. The model is designed for token efficiency, which may lower costs compared to routing the same tasks to frontier models like Claude or ChatGPT.
Salesforce's Agentforce platform will offer Koa alongside other available models, with an AI gateway determining which model handles each request based on the task. The company also announced a separate partnership with Anthropic called ClaudeForce, showing that Salesforce is not abandoning frontier providers entirely. Salesforce executives noted that before Nemotron, no suitable American pre-trained base model with clear data provenance was available for this kind of enterprise-grade post-training work.
Koa may shift enterprise AI adoption by offering a more cost-effective, privacy-preserving alternative to closed frontier models. Businesses using Agentforce could reduce AI spending while maintaining data security, potentially making AI agents more accessible to smaller enterprises. However, the model's reliance on synthetic data and specific task focus may limit its versatility, and the broader impact on AI competition remains uncertain as companies weigh open-weight versus proprietary approaches.