Salesforce launches Koa, an enterprise reasoning model post-trained with NVIDIA on 27 years of workflows
silviocinguetta · x · 2026-09-18
At Dreamforce 26, Salesforce announced Koa, its first reasoning model optimized for enterprise work, co-engineered with NVIDIA by post-training Nemotron 3 Super, running inside the Salesforce trust boundary to power Agentforce agents.
Key arguments from EVPs Jayesh Govindarajan and Silvio Savarese:
- Frontier models reason about your business "from the outside," treating a refund policy like a physics problem.
- Over 18 months Salesforce split intent detection, prompt-injection screening, toxicity, evals and reranking into small dedicated models, but left multi-step reasoning to a general model — until now.
- Koa is like employee onboarding: post-trained on 27 years of enterprise workflows to internalize compliance lines, credit thresholds and lead qualification.
A representative case of the domain-specific post-trained reasoning model approach.
More from Companies & People
- Hard Fork signs off: Kevin Roose and Casey Newton launch Machine Gods with NPR — kevinroose · 2026-09-18
- Waymo Announces Its Second Asia City for Robotaxi Expansion — reed · 2026-09-18
- Can AI-native sales stacks disrupt Salesforce's $200B? Owning the sale, not the CRM — CCgong · 2026-09-18
- Mistral customer data leak sparks doubt over 'open models are private' myth — Afinetheorem · 2026-09-18
- Foresight's Vision Weekend to gather Neuralink, Science Corp, Paradromics on whole brain emulation — irinarish · 2026-09-18
- Notion CEO Ivan Zhao on Engelbart, Lego, and building tools with an artist's care — ivanhzhao · 2026-09-18