Salesforce unveils Koa, its first reasoning model built on Nvidia’s Nemotron

Salesforce just entered the custom reasoning model race with Koa, a purpose-built AI system derived from Nvidia’s Nemotron-3-Super-120B foundation model. The model is designed to handle multi-turn reasoning and tool-use tasks inside Salesforce’s Agentforce platform, with a laser focus on the kind of work that actually pays the bills: sales pipelines, marketing campaigns, and customer support tickets. What Koa actually does Koa is not a general-purpose chatbot. It’s a specialized reasoning model trained to navigate the complex, multi-step workflows that enterprise CRM systems demand. The model was trained exclusively on public and synthetically generated data. The training process uses a reinforcement learning technique called Group Relative Policy Optimization, or GRPO, which incorporates what Salesforce describes as a simulation-to-reward pipeline. In practical terms, the system learns by simulating real Agentforce workflows (called Agent Scripts) and optimizing for outcomes that mirror successful task completion. On benchmarks, Koa outperforms the Nemotron base model it was derived from, particularly in multi-turn tool-use scenarios. It also beats strong proprietary baselines acr...