Salesforce and Nvidia launch Koa reasoning model

At its Dreamforce conference on September 15, 2026, Salesforce announced Koa, a specialized AI reasoning model developed in partnership with Nvidia. Built on Nvidia's open-weight Nemotron architecture, Koa is designed to execute complex, multistep CRM workflows within Salesforce's Agentforce platform. To ensure data privacy, the model was trained entirely on synthetic data simulating 30 years of enterprise workflows. Koa matches or exceeds general-purpose models on CRM tasks with three times fewer errors, offering enterprises a highly secure, token-efficient alternative to frontier models.

Koa reasoning model launch

• ▪The Koa model is currently available to select customers via an expanded pilot program, with general availability slated for winter
• ▪Salesforce Inc. announced the Koa model at its annual Dreamforce tech conference on September 15, 2026
• ▪Salesforce Inc. partnered with Nvidia Corp. to train and release Koa, a specialized artificial intelligence model custom-built for reasoning about customer relationship management data

Nvidia Nemotron base model

• ▪Salesforce Inc. will integrate Nvidia's full suite of Nemotron models within its secure Missionforce Operations platform starting in October 2026
• ▪Salesforce Inc. selected Nvidia's Nemotron as a base model because it provided a sovereign American, state-of-the-art pre-trained model with clear data provenance
• ▪The Koa model is built on Nvidia Corp.'s open-weight Nemotron model architecture

Synthetic data training approach

• ▪Salesforce Inc. and Nvidia Corp. trained Koa without using any actual customer data, instead utilizing a synthetic dataset simulating realistic enterprise workflows
• ▪Salesforce Inc. generated its synthetic training dataset by leveraging nearly 30 years of internal customer relationship management deployment experiences across 14 industries
• ▪The synthetic training process for Koa combined supervised fine-tuning, reinforcement learning, and group relative policy optimization to teach the model multistep tool execution

Enterprise AI divergence trends

• ▪The development of specialized models like Koa highlights a divergence where enterprise AI needs favor open-weight, task-specific models over general-purpose frontier models
• ▪Salesforce Inc. announced a partnership with Anthropic PBC called Claudeforce, allowing customers to use Claude as their AI interface while keeping data secured within Salesforce

Agentforce platform integration

• ▪Prior to Koa, Salesforce's Agentforce platform relied on routing complex reasoning tasks to frontier models like Anthropic's Claude or OpenAI's ChatGPT
• ▪Koa is hosted and run within Salesforce's trust boundaries via Nvidia's open framework, ensuring customer data does not leave the controlled environment during inference
• ▪Koa is designed to help AI agents reason through complex, multistep workflows within Salesforce's Agentforce platform

Token efficiency cost reduction

• ▪Koa is designed to reduce enterprise AI spending by using fewer tokens to execute the same work tasks compared to general-purpose frontier models
• ▪In benchmark tests, Koa matched or exceeded the performance of multiple general-purpose models on CRM tasks while producing three times fewer errors

Debatable claims

• ▪Synthetic data is sufficient for training enterprise-grade AI reasoning models
• ▪Enterprises should prioritize task-specific AI models over general-purpose frontier models