Salesforce Debuts Koa, a CRM Reasoning Model Built on NVIDIA Nemotron

Salesforce introduced Koa on September 15, 2026, during Dreamforce week. It is the company's first reasoning model built specifically for CRM work, presented in San Francisco by Salesforce CEO Marc Benioff and NVIDIA founder and CEO Jensen Huang.

How Koa was trained

Koa was created by post-training NVIDIA's open Nemotron 3 Super model. The training corpus is synthetic and modeled on enterprise knowledge from nearly three decades of Salesforce CRM deployments; the company says no customer data was used at any stage. The scenarios span more than 14 industries, including manufacturing, financial services, healthcare and travel. Each one pairs a persona with specific tasks and maps the sequence of actions and tool calls an agent should take.

Post-training combined supervised fine-tuning with reinforcement learning using Group Relative Policy Optimization (GRPO), carried out with NVIDIA's NeMo RL, NeMo Gym and NeMo AutoModel tools. Because Nemotron is open, Salesforce controls the model weights and runs both training and inference inside its own infrastructure.

Benchmarks and first customers

On Salesforce's CRM benchmark, which includes real-world tasks such as updating an opportunity, routing a case or scheduling a follow-up, Koa matches or exceeds leading models while producing three times fewer errors on those actions, according to the company. Koa already runs inside Salesforce, where it powers an employee agent in Slack. In October it moves into customer pilots as a customer-selectable model in Agentforce, with an initial group that includes Formula 1, UChicago Medicine, Baxter Credit Union, 1-800Accountant, Engine and Xero. General availability is expected in winter 2026 for U.S. regions.

Missionforce and the open-model shift

Salesforce and NVIDIA are also bringing Nemotron-based models and accelerated computing into Missionforce, aimed at government and regulated organizations that need to run AI in private clouds, classified networks or fully air-gapped environments. Post-trained NVIDIA models will reach select customers in October. Huang said open-model adoption among enterprises building custom AI systems has grown from roughly 30% to about 70%. He also argued that AI safety is an engineering problem rather than a reason to slow down, telling Benioff that companies should run as fast as they can but pause if a product is not ready.