On September 15, 2026, Salesforce and NVIDIA announced Koa, Salesforce’s first CRM reasoning model for Agentforce, built on NVIDIA Nemotron. Koa was built by post-training NVIDIA Nemotron 3 Super with a proprietary synthetic dataset modeled on nearly three decades / ~27 years of CRM deployments. Salesforce CRM benchmark: matches or exceeds leading model performance on CRM actions with 3x fewer errors (Salesforce claim). Salesforce controls model weights; post-training and inference within Salesforce trust boundary. No customer data used to train Koa; synthetic scenarios across 14+ industries; SFT + GRPO with NVIDIA NeMo RL, NeMo Gym, NeMo AutoModel. Missionforce: Nemotron-based models + accelerated computing for government/regulated; private clouds, air-gapped. Short quotes: Marc Benioff (Salesforce), Jensen Huang (NVIDIA). Pilots: 1-800Accountant, Baxter Credit Union (BCU), Engine, Formula 1, UChicago Medicine, Xero; also internal Slack agent. Availability: Koa select pilots now in Agentforce; GA expected winter 2026 U.S. regions. Missionforce Operations GA now U.S.; post-trained NVIDIA models to select customers October 2026.
Salesforce is putting CRM-specific reasoning into a model hosted inside its trust boundary, with NVIDIA Nemotron as the base and Missionforce for regulated/air-gapped deployments. Track as specialized CRM reasoning + Agentforce packaging — keep 3x fewer errors as a Salesforce claim; treat pilot/GA timing as Salesforce statements.
Pro sample · free to use
Connect the whole task.
Try the meeting follow-through pack: agenda, action register, follow-up and weekly update. The sample is available now; the paid pack library is planned.
Pro proposal: US$7/month or US$70/year. No card or payment today.
- 01
Sep 15, 2026: Salesforce + NVIDIA announce Koa — first CRM reasoning model for Agentforce on NVIDIA Nemotron.
- 02
Base: Post-trained NVIDIA Nemotron 3 Super on proprietary synthetic CRM dataset (~27 years / nearly three decades framing).
- 03
Note: CRM benchmark matches/exceeds leading models with 3x fewer errors = Salesforce claim.
- 04
Trust: Salesforce controls weights; post-training/inference in trust boundary; no customer data used to train (Salesforce).
- 05
Stack: SFT + GRPO with NeMo RL / Gym / AutoModel; 14+ industry synthetic scenarios.
- 06
Missionforce: Nemotron + accelerated computing for gov/regulated; private/air-gapped; Operations GA now U.S.; post-trained models select customers Oct 2026.
- 07
Availability: Select pilots now (named customers + Slack agent); Koa GA expected winter 2026 U.S. regions.
Editorial text reviewed 18 September 2026. Publisher claims have not been independently retested by ElevateU.
Last reviewed: 2026-09-17T13:20:00Z · Speaker predictions remain attributed to the speaker. ElevateU analysis is labeled separately from source-supported fact.