Key facts
- Salesforce launched Koa, its first AI reasoning model, at its Dreamforce conference.
- Koa is built on Nvidia's open-weight Nemotron model.
- The model is designed for sales, marketing, and customer-support tasks.
Salesforce unveiled Koa, its first AI reasoning model, developed in partnership with Nvidia using its Nemotron model. Koa is designed for specific enterprise tasks in sales, marketing, and customer support, offering an open-weight alternative to proprietary models and aiming to reduce AI spending.

The launch of Koa signifies a shift in enterprise AI strategy, offering businesses more control over their data and AI costs by providing a specialized, open-weight model that competes with proprietary offerings from major AI labs.
Salesforce has introduced Koa, its inaugural AI reasoning model, developed in collaboration with Nvidia. Unveiled at the company's Dreamforce conference, Koa is built upon Nvidia's open-weight Nemotron model and has been specifically post-trained to excel in sales, marketing, and customer support functions. This initiative represents a divergence from the offerings of frontier AI labs, which often require enterprises to upload sensitive data into their proprietary systems.
Koa provides Salesforce customers with an open-weight alternative to closed models, a tool tailored for specific work tasks rather than general problem-solving, and a model that does not ingest customer data, thereby preventing potential leaks. It is also designed to reduce AI spending by using fewer tokens for equivalent tasks and can be integrated with Salesforce's existing data security and routing protocols. Jayesh Govindarajan, EVP of Salesforce AI, stated that while Salesforce had previously developed smaller task-specific models, reasoning capabilities were always outsourced to frontier models until the advent of Koa.
Previously, complex or multi-step tasks requiring reasoning were routed through Salesforce's AI gateway to models like Claude or ChatGPT. Govindarajan highlighted the lack of a suitable pre-trained base model as a previous barrier to developing their own enterprise-grade model, noting that Nvidia's Nemotron provided a state-of-the-art, sovereign American option with clear data provenance. The post-training process involved crafting synthetic data that mimicked customer interactions to refine Koa's knowledge in sales and customer support.
Kari Ann Briski, Nvidia's VP of Generative AI Software for Enterprise, emphasized the unique architecture of Nemotron for token efficiency, calling it a "trifecta" of sovereign AI, fast inference, and efficient tokenomics. Despite the development of Koa, Salesforce is continuing its relationship with Anthropic, announcing a ClaudeForce partnership that allows companies to use Claude as an AI interface while keeping their data secured within Salesforce's infrastructure.