Salesforce says its Koa model matches leading models on CRM actions with three times fewer errors.
The weights are Nvidia's open Nemotron, the post-training is Salesforce's, and Salesforce keeps both the weights and the inference in house.
Salesforce and Nvidia announced a model called Koa on September 15, 2026, from San Francisco. It's Salesforce's first CRM reasoning model, and it was built by post-training Nvidia's Nemotron 3 Super.
No customer data went into it. Salesforce says the training corpus was synthetic from end to end, written to simulate enterprise workflows across more than 14 industries, and that it holds the model weights and runs inference inside its own trust boundary.
- Model
- Koa, Salesforce's first CRM reasoning model
- Built from
- Nvidia Nemotron 3 Super, post-trained by Salesforce
- Training data
- synthetic scenarios, more than 14 industries, no customer data
- The claim
- three times fewer errors on Salesforce's CRM benchmark
- Availability
- select pilot customers now, general availability expected winter 2026
What they trained it on
Each synthetic scenario paired a persona with specific tasks, then mapped the sequence of actions and tool calls an agent has to make to finish them. Salesforce post-trained with supervised fine-tuning and Group Relative Policy Optimization, using Nvidia's NeMo RL, NeMo Gym and NeMo AutoModel.
Salesforce says that taught the model to take the right action step by step, and not only to produce the right answer. That's the gap between a chat window and something that updates a record in your system.
The performance claim is narrow and specific. On Salesforce's own CRM benchmark, which it describes as real-world tasks like updating an opportunity, routing a case, or scheduling a follow-up, Koa "already matches or exceeds leading model performance on CRM actions with three times fewer errors". Salesforce doesn't say which leading models it measured against.
With Koa, the knowledge is put inside the model itself.
Salesforce didn't buy this from a frontier lab. It took an open-weight model, spent its post-training on one domain, and came out claiming parity on the work its customers do all day. The release doesn't name a frontier lab anywhere in it.
Marc Benioff, Salesforce's chair and CEO, said the company trained "a reasoning engine that understands the structure of a deal, the lifecycle of a service case, and the workflows that vary across industries". His argument is that the accumulated knowledge of how enterprise business works, rather than the platform, is what Salesforce has that's hard to copy.
What Nvidia gets
NVIDIA Nemotron open models give Salesforce the foundation to turn decades of enterprise expertise into specialized AI with Koa, creating a CRM model that can reason and securely take action.
Jensen Huang, Nvidia's founder and CEO, frames the same deal as a larger opportunity for software. It also brings Nemotron-based models and Nvidia computing into Missionforce, the Salesforce product for government and regulated customers, which has to run in private clouds, classified networks and fully air-gapped systems.
Who has it
Koa is already running inside Salesforce, including an agent in Slack that helps employees find information and get everyday tasks done. Customer pilots have started at 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.
Salesforce expects general availability in U.S. regions in winter 2026. Post-trained Nvidia models reach select Missionforce customers in October 2026.