Salesforce Unveils In-House AI Model Koa Built on Nvidia's Open-Weight Nemotron

Salesforce has developed its first in-house AI model, Koa, using Nvidia's open-weight model family, Nemotron. The company built Koa by post-training Nemotron with synthetic scenarios that mimic customer interactions, ensuring the model doesn't leak sensitive data. This approach allows Salesforce to control and audit the model without relying on pre-trained general-purpose models.

Nemotron 3 Ultra, the basis for Koa, is a large language model with 550 billion parameters. While it ranks mid-table in the Artificial Analysis Intelligence Index, it outperforms its Chinese counterparts in speed and cost. Salesforce claims Koa achieves three times fewer errors than leading models on CRM actions and offers 11 percent more precision in calling the right action.

Koa runs entirely within the Salesforce trust boundary, ensuring security and consistency. The model is available to selected pilot customers now, with general availability announced for winter regions. In parallel, Salesforce is introducing Nemotron 3 Nano into Agentforce, a context window of one million tokens, and the option to run it on premises or in a private cloud.