{"id":406,"date":"2026-09-16T07:40:08","date_gmt":"2026-09-15T23:40:08","guid":{"rendered":"https:\/\/aidashxp.com\/salesforce-koa-reasoning-model\/"},"modified":"2026-09-16T07:40:08","modified_gmt":"2026-09-15T23:40:08","slug":"salesforce-koa-reasoning-model","status":"publish","type":"post","link":"https:\/\/aidashxp.com\/en\/salesforce-koa-reasoning-model\/","title":{"rendered":"Salesforce Launches Its First Reasoning Model, Koa: Built with Nvidia for Enterprise-Specific AI"},"content":{"rendered":"

At this week\u2019s Dreamforce conference, Salesforce unveiled its first true reasoning model\u2014Koa<\/strong>Built on Nvidia\u2019s open-source weights Nemotron<\/strong> This enterprise-specific model, trained on synthetic data, targets sales, marketing, and customer support scenarios\u2014and TechCrunch dubbed it \u201csomething that should scare AI labs.\u201d<\/p>\n\n\n\n

What is Koa<\/h2>\n\n\n\n

Koa is a reasoning model jointly post-trained by Salesforce and Nvidia. It refines the open-source Nemotron base with synthetic data to become a \u201cbusiness expert\u201d fluent in sales dialogue, customer service de-escalation, and marketing copy. Jayesh Govindarajan, Executive Vice President of Salesforce AI, stated plainly:\u201cFor reasoning, we\u2019ve historically relied on frontier model providers\u2014no longer.\u201d<\/strong><\/p>\n\n\n\n

Choosing Nemotron as the base was no coincidence. Govindarajan explained that before Nemotron, there was no U.S.-origin open-source pre-trained model with both transparent data provenance and frontier-level performance\u2014\u201cWe had no idea what data Qwen (Alibaba\u2019s open-source model) was trained on.\u201d With an open-source base featuring traceable data sources, enterprises gain the confidence to conduct post-training safely.<\/p>\n\n\n\n

Why enterprises will pay<\/h2>\n\n\n\n