Salesforce and Nvidia introduce Koa, a task-specific reasoning model

Salesforce has announced Koa, its first internal reasoning model, developed in collaboration with Nvidia at the Dreamforce tech conference. Built on top of Nvidia’s open-weight Nemotron model, Koa has been post-trained specifically for enterprise tasks in sales, marketing, and customer support.

The release highlights a growing divergence between the demands of enterprise software users and the offerings of consumer-facing frontier AI laboratories. While major labs often encourage businesses to upload proprietary data, code, and prompts directly into cloud environments, enterprise buyers frequently require strict security boundaries, sovereign base models, and predictable cost structures.

Core architectural advantages

Salesforce plans to integrate Koa into its Agentforce platform as an alternative to external models. The system introduces several distinct characteristics for enterprise environments:

• Built on an open-weight alternative to closed frontier systems.
• Tuned specifically for commercial workflows rather than general academic benchmarks.
• Trained entirely without actual customer data to eliminate leakage risks.
• Optimized for lower token usage to reduce operational expenditures.
• Integrated directly into existing data security infrastructure.

Training with synthetic environments

To bridge the gap between general-purpose foundations and business utility, the development teams avoided using real customer logs during post-training. Instead, they relied on synthetic data generation to simulate realistic workplace scenarios.

Simulations included various operational dynamics, ranging from routine scheduling to managing difficult interactions with simulated irate customers. This approach allows the model to develop conversational and problem-solving patterns tailored to commercial settings.

Despite the introduction of Koa, Salesforce maintains broader industry partnerships. The company simultaneously announced a collaboration allowing enterprises to utilize external systems such as Anthropic’s Claude within secured infrastructure frameworks.