Salesforce launches inference model Koa: betting on open weights and synthetic data to protect customer privacy

Salesforce announced the launch of the Koa model, the first inference model dedicated to customer relationship management (CRM) systems, built on Nvidia's open Nemotron 3 Super model. The announcement came at the Dreamforce conference, where Nvidia CEO Jensen Huang and Salesforce CEO Marc Benioff shared the stage, confirming that artificial intelligence is becoming a global infrastructure layer that runs enterprise operations through specialized digital agents.

The model's development relied on fine-tuning the Nemotron model with synthetic data derived from three decades of experience deploying CRM solutions across more than 14 sectors, including financial services, manufacturing, health care and travel.Salesforce said that training and inference of the model did not consume a single byte of actual customer data.The open license for Nvidia's model allowed full control over the weights and their complete operation within Salesforce's private infrastructure, using the NeMo RL, NeMo Gym and NeMo AutoModel training and reinforcement-learning tools.

Performance test results on the enterprise CRM Bench benchmark, which measures real-world tasks such as updating sales opportunities, routing tickets and scheduling follow-ups, showed that the Koa model matches or exceeds leading customer-management models while recording three times fewer errors. Salesforce began running the model internally to support employee agents via the Slack platform, with pilot trials for subscribers slated to start in October on the Agentforce platform for clients such as Formula 1, Chicago Medicine and Bexar Credit Union, paving the way for a general release in the U.S. during the winter of 2026.

Huang explained that the share of reliance on open models jumped from 30 percent at the start of last year to about 70 percent now, noting that organizations are simultaneously moving toward building their own models so that every software company becomes an AI-agent company. Huang rejected arguments that the technology threatens human jobs or the software industry, emphasizing that "AI safety is a purely engineering issue, not a choice that conflicts with the pace of innovation, the fundamental standard is to refrain from launching any product whose capabilities or safety are untrusted and to take time to address any flaws."

This shift has direct implications for those leading digital transformation and cloud-infrastructure management in the Gulf, Egypt and the region, especially in banking, telecommunications and government agencies that operate under strict data-governance frameworks whose sovereignty prevents leaking customer records to closed-model interfaces. The Koa model demonstrates that it is possible to develop superior functional inference without compromising sensitive data by relying on synthetic data and open weights, giving technology leaders a genuine alternative that reduces the cost of calling external APIs and spares them regulatory risk.

Your practical step as a technology officer is to review the efficiency of internal customer-service and sales agents and assess whether relying on specialized, synthetically trained open inference models offers higher accuracy and lower operating cost compared with general-purpose multitask models.