Salesforce Koa

Salesforce's first branded language model: an enterprise agentic model built by post-training NVIDIA's open-weight Nemotron-3-Super-120B with supervised fine-tuning and GRPO reinforcement learning on public and synthetic data only, with no customer data. The report's distinctive piece is a simulation-to-reward pipeline that expands workflow specifications, written in Agent Script for enterprise domains, into persona-conditioned multi-turn tasks whose rewards are grounded in successful tool use for data-dependent requests. Reported results: τ²-Bench 69.41, BFCL 66.63%, and 0.86 on Salesforce's CRM Bench against Claude Opus 4.8's 0.87, with "3x fewer errors" on CRM actions. Weights are closed; Koa runs inside Agentforce, in pilot from September 2026 with general availability planned for winter 2026.