Salesforce and NVIDIA ship Koa, a CRM-native reasoning…

Post-trained on Nemotron 3 Super with synthetic CRM data, Koa enters pilot today; GA is targeted for winter 2026.

Salesforce and NVIDIA unveiled Koa today at Dreamforce in San Francisco, a CRM-native reasoning model post-trained on NVIDIA’s open-weight Nemotron-3-Super-120B and pointed straight at the multi-step work that Agentforce agents have been outsourcing to frontier labs. Pilot begins now with six named customers: 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero. General availability is targeted for U.S. regions in winter 2026.

The framing from Salesforce EVP of AI Jayesh Govindarajan is the tell. “Reasoning has always been something that we’ve relied on the frontier model providers for. Until now.” What Salesforce wanted, he said, was “a sovereign American pre-trained model that was available… state of the art, and… had clear data provenance.”

Translation: the AI gateway that has been quietly routing lead qualification and opportunity updates to Claude and ChatGPT is now a cost center, and Salesforce would prefer to own the reasoning layer inside its own trust boundary.

Koa was trained on synthetic scenarios modeled on nearly three decades of CRM deployments across more than 14 industries, with no customer data used. A September 14 arXiv paper describes a simulation-to-reward pipeline using Group Relative Policy Optimization, with Agent Script driving persona-conditioned tasks through NeMo RL, NeMo Gym, and NeMo AutoModel. Salesforce claims three times fewer errors than leading models on CRM actions. The paper itself is more measured: gains over the open-weight base are strongest on multi-turn tool use, and Koa surpasses a strong proprietary baseline while remaining below the strongest frontier models. Moor Insights & Strategy notes that Salesforce is grading its own test.

NVIDIA’s Kari Ann Briski described Nemotron as “token efficient” and pitched the bundle plainly: “It’s kind of the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all.”

Koa sits alongside the Claudeforce interface tier and the wider Agentforce agent roster, with Nemotron-based models reaching Missionforce Operations customers in October 2026. The model wars have a new front, and it’s inside the CRM.

Salesforce and NVIDIA ship Koa, a CRM-native reasoning model for Agentforce

Sources