Overview
- Koa, which Salesforce unveiled Tuesday at Dreamforce, is a reasoning model post-trained from Nvidia’s Nemotron-3-Super-120B to handle multi-turn CRM tasks such as deal progression and escalated support cases.
- Salesforce and Nvidia say Koa was trained only on public and synthetically generated data using a reinforcement learning method called Group Relative Policy Optimization (GRPO) to simulate agent-and-customer workflows.
- Nemotron’s architecture gives Koa a 1 million‑token context window and a Mixture-of-Experts design that Nvidia says lowers inference token use and speeds time to first token for long, tool-driven conversations.
- Koa will be offered inside Salesforce’s Agentforce as an on-platform alternative to routed frontier models like Claude, while Salesforce keeps integrations so customers can choose models based on cost, capability, and compliance.
- Salesforce reports Koa beats its Nemotron base and strong proprietary baselines on CRM multi-turn benchmarks but still trails the largest frontier models, leaving independent pilots, cost comparisons, and third‑party benchmarks as key next steps.