Salesforce's first reasoning model Koa released, based on Nvidia Nemotron 120B

动察 Beating AI News Flash: Salesforce has teamed up with Nvidia to launch Koa, its first reasoning model designed for CRM. It is post-trained on the open-source Nemotron-3-Super-120B and is specifically built to handle multi-step tasks in sales, marketing, and customer service. Previously, Agentforce's reasoning relied mainly on external models such as OpenAI and Anthropic.

Koa was not trained using Salesforce customer data. The team used public data and synthetic scenarios to simulate CRM workflows, then trained the model through reinforcement learning to call tools and complete tasks in sequence. Koa outperforms GPT-4.1 on all three tests—Tau2Bench, BFCL, and CRM Bench—but trails GPT-5.5 and Claude Opus 4.8 on all three.

Koa will not replace external models, but will become another model option for Agentforce, coexisting with models from OpenAI, Anthropic, and others. Salesforce can assign tasks better suited to CRM to its own model, while continuing to call third-party models for the rest.