Salesforce Owns a CRM Model — Koa on Nemotron, Weights Inside the Trust Boundary
Salesforce and NVIDIA announced Koa, a CRM reasoning model post-trained from Nemotron 3 Super on synthetic data reflecting 27 years of CRM experience. Salesforce reports 3× fewer CRM Benchmark errors. Pilots are underway; US general availability is planned for winter 2026.
Correction / update, October 14, 2026: Salesforce’s no-customer-data statement concerns training. Inference is described as staying within its trust boundary, not as using no customer data.
Enterprises are post-training American open weights instead of sending CRM actions to Claude or GPT. Salesforce is owning a CRM model — and keeping the weights inside its trust boundary.
Salesforce and NVIDIA at Dreamforce (September 15) announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset modeled on ~27 years of CRM deployments. No customer data was used in training, Salesforce said; it controls the weights and runs post-training and inference inside its own trust boundary. Vendor CRM Benchmark (update opportunity, route a case, schedule a follow-up): Koa matches or exceeds leading general-purpose models with 3× fewer errors. Already running an employee Slack agent inside Salesforce. Customer pilots: Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine, Xero. Pilot now; GA winter 2026, US regions. Separate: Nemotron models into Missionforce (gov / air-gap); post-trained SKUs for select customers in October 2026.
Koa is a post-trained CRM model, separate from the Salesforce in Claude plugin and the underlying Nemotron open weights.
What shipped#
| Item | Status |
|---|---|
| Koa in Agentforce | Select pilots now; GA winter 2026 US |
| Training | SFT + GRPO via NeMo RL / Gym / AutoModel; 14+ industries, synthetic only |
| Missionforce Operations | GA now (US); post-trained NVIDIA models October |
Claims vs checks#
The cited results are from Salesforce’s CRM Benchmark; this article has not verified an independent CRM benchmark or leaderboard result. Treat 3× fewer errors as vendor, not independent. Watch whether that number survives a third-party CRM bench, and whether winter GA stays US-only.