Salesforce ships Koa, an enterprise agent model post-trained on Nvidia's open-weight Nemotron 3 Super

TechCrunch (corroborated by arXiv 2609.15066)·high signalXBlueskyLinkedInCopy link

Salesforce built Koa by post-training the open-weight Nemotron-3-Super-120B (120B total, 12B active, hybrid Mamba-Transformer MoE) with GRPO reinforcement learning on public and synthetic data only, no customer data. The paper (arXiv 2609.15066) reports Koa beating its open-weight base most clearly on multi-turn tool use and surpassing GPT-4.1 on public tool-use, agentic-reasoning and CRM benchmarks while staying below frontier models. The builder takeaway is that a vendor can now take an open 120B base and RL it into a competitive domain agent without a frontier pretraining run.