From Mexico — Salesforce and NVIDIA just announced Koa: Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Same-day coverage from TechCrunch at Dreamforce.
I’m reading this as an enterprise-specialization story, not another general chat model drop. Salesforce says it post-trained Nemotron 3 Super with a proprietary synthetic dataset modeled on nearly three decades of CRM deployments — and that no customer data was used to train Koa.
What Salesforce says Koa does
Per the company story, Koa is meant to help agents reason through complex, multistep CRM workflows and pick the right tools — things like updating an opportunity, routing a case, or scheduling a follow-up. Salesforce controls the model weights and says post-training and inference stay inside its own trust boundary.
On Salesforce’s own CRM benchmark, the company claims Koa already matches or exceeds leading-model performance on CRM actions with three times fewer errors. That’s Salesforce’s scoreboard — I’m not inventing third-party bake-offs.

Why Nemotron mattered (TechCrunch)
TechCrunch quotes Jayesh Govindarajan, Salesforce’s EVP of AI: reasoning used to get routed to frontier providers through Agentforce’s AI gateway. The blocker for training their own enterprise-grade reasoning model, he said, was lacking a suitable pre-trained base — until Nemotron — with clear data provenance.
Salesforce isn’t dumping Anthropic or OpenAI either. TechCrunch notes a new Anthropic partnership branded ClaudeForce: Claude as the AI interface while customer data stays in Salesforce’s system of record.
Availability — pilots, not full GA
Salesforce: Koa is available to select pilot customers now in Agentforce; general availability expected winter 2026 in U.S. regions. Named pilot logos in the story include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine, and Xero — I’m listing them as the company does, not as proof of production scale.
My takeaway
From Mexico, the narrow read: Salesforce × NVIDIA shipped a CRM-specialized reasoning model for Agentforce on Nemotron 3 Super, trained on synthetic enterprise scenarios (company says no customer data), with Salesforce-hosted weights and pilots now — U.S. GA still winter 2026. Sources: Salesforce, TechCrunch.