Salesforce & Nvidia launch Koa, an enterprise reasoning model
Original: Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
Why This Matters
Koa shows enterprises can build competitive reasoning models without frontier labs — threatening OpenAI and Anthropic's enterprise revenue.
Salesforce unveiled Koa at Dreamforce 2026, its first reasoning model built on Nvidia's open-weight Nemotron. Post-trained on synthetic data for sales, marketing, and customer support, Koa offers enterprises an alternative to closed frontier models like Claude or ChatGPT — with lower token costs and no customer data ingestion.
Koa is Salesforce's debut reasoning model, announced at Dreamforce 2026 and built atop Nvidia's open-weight Nemotron. The two companies jointly post-trained the model using synthetic data — not actual customer records — to specialize it for sales, marketing, and customer support workflows.
Before Koa, when an Agentforce task required multi-step reasoning, Salesforce routed those prompts to frontier models like Claude or ChatGPT. That's now changing. Salesforce EVP of AI Jayesh Govindarajan told TechCrunch: 'Reasoning has always been something that we've relied on the frontier model providers for. Until now.'
Govindarajan cited Nemotron's clear data provenance as a key differentiator, taking a direct shot at Alibaba's Qwen: 'We have no idea what Qwen trains on.' He added that until Nemotron, no sovereign American open-weight base model was both state-of-the-art and transparently sourced.
Koa delivers six enterprise-specific advantages: it's open-weight, task-specialized, free of customer data, token-efficient, gatewayroutable via Agentforce, and governed by Salesforce's existing security and compliance controls. It will sit alongside other models in Agentforce's model selection system, letting the platform auto-route requests depending on task complexity.