Salesforce

Enterprise Software · Artificial Intelligence · CRM

Launch

Salesforce and Nvidia launch Koa, a CRM-specific AI reasoning model

Built on open-weight Nemotron rather than a closed frontier model, it targets a gap the big AI labs have largely ignored.

• Salesforce and Nvidia announced Koa at Dreamforce on September 15, 2026, Salesforce's first CRM reasoning model for its Agentforce platform, built on Nvidia's open-weight Nemotron 3 Super.
• Koa was created by post-training Nemotron on a synthetic dataset modeled on 27 years of Salesforce CRM deployment data spanning 14+ industries like manufacturing, healthcare, and financial services.
• On Salesforce's CRM Benchmark—tasks like updating an opportunity, routing a case, or scheduling a follow-up—Koa matches or exceeds leading models while making three times fewer errors.
• Salesforce controls the model weights and runs all post-training and inference within its own infrastructure, so no customer data leaves its trust boundary to train the model.
• Before Koa, Agentforce routed complex multi-step reasoning tasks to frontier models like Claude or ChatGPT through its AI gateway; Salesforce's AI leadership said a sovereign, pre-trained base model wasn't available until Nemotron.
• Salesforce and Nvidia are also bringing Nemotron-based models and accelerated computing into Missionforce, extending model and data control to government and regulated organizations in private clouds and air-gapped networks.
• Koa signals enterprise AI splitting from general-purpose frontier models: task-specific, cheaper-to-run models trained on proprietary domain data may become the norm for repetitive business workflows rather than sending everything to closed labs.