‘Now We Can Know Everything and Do Anything,’ Jensen Huang Says at Dreamforce
The CRM reasoning model was post-trained on synthetic enterprise data using NVIDIA NeMo tools and enters customer pilots in October.
- Salesforce built Koa by fine-tuning and applying reinforcement learning to NVIDIA Nemotron 3 Super using NeMo RL, NeMo Gym, and NeMo AutoModel
- The training dataset uses synthetic enterprise scenarios spanning over 14 industries and no customer data
- Salesforce states Koa matches or exceeds leading models on its CRM Bench evaluation suite with 3x fewer errors
- Koa is currently running internally to power a Slack employee agent, with general availability scheduled for winter 2026 in US regions
Salesforce customers get a domain-specialized reasoning model built entirely on self-hosted open model weights without exposing enterprise CRM data to external APIs.
