Salesforce launches Koa, its first CRM reasoning model, built on NVIDIA Nemotron 3 Super

Salesforce and NVIDIA used Dreamforce in San Francisco to announce Koa, which Salesforce calls its first CRM reasoning model for Agentforce. The model was built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset drawn from nearly three decades of enterprise CRM deployments.

NVIDIA founder and CEO Jensen Huang joined Salesforce CEO Marc Benioff onstage at the Moscone Center and walked through the audience with a microphone in hand. "With electricity, we could power everything. With the internet, you can find anything. And now, with artificial intelligence as an infrastructure layer across the planet, we can know everything and do anything," Huang said, framing AI as the next general-purpose infrastructure layer.

Koa was trained with supervised fine-tuning and reinforcement learning using NVIDIA NeMo RL, NeMo Gym, and NeMo AutoModel, and its training corpus covers synthetic enterprise scenarios across more than 14 industries, including manufacturing, financial services, healthcare, and travel. Because NVIDIA Nemotron is open, Salesforce was able to fine-tune and run the model entirely within its own infrastructure, controlling the weights. "Not a single byte of customer data was used," said Rohan Kumar, Salesforce's president of platform and engineering.

Salesforce says that in its own CRM Bench, a benchmark of real-world tasks such as updating an opportunity, routing a case, or scheduling a follow-up, Koa already matches or exceeds leading model performance on CRM actions while making three times fewer errors.

Koa is already running inside Salesforce, powering an employee agent in Slack, and it moves into customer pilots in October as a customer-selectable model in Agentforce. The first named pilot customers are Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine, and Xero. General availability is expected in winter 2026 in U.S. regions.

Asked about AI safety, Huang called it job one but framed it as an engineering problem: if you build a product or service and you are not confident in its functionality, capability, or safety, then don't release it. He rejected the idea that speed and safety are a trade-off, telling companies to run as fast as they can but to pause and get things right if a product is not going to be safe.

Huang also said open models have grown from roughly 30 percent of usage at the start of last year to about 70 percent today, arguing that companies are adopting closed models at exponential rates while also building their own custom AI, because every software company is an AI company and every enterprise is becoming an agentic one. NVIDIA is a case in point, running Salesforce across sales, service, marketing, and operations and piloting Agentforce for customer-support workflows.

The announcement matters because CRM vendors are starting to bake their own domain data into the model layer instead of wrapping a general-purpose model in workflow software. The open-weight approach also gives Salesforce a direct answer to the compliance concern that stalls so many enterprise AI deployments. What to watch next: how the October pilots report back, whether third-party benchmarks confirm the CRM Bench numbers, and how far Koa spreads beyond U.S. regions after general availability.