Tech news in 3 minutes

Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear

6 h ago

Salesforce unveiled Koa, its first AI reasoning model built on Nvidia’s open-weight Nemotron, at Dreamforce this week, marking a major enterprise AI shift toward task-specific, cost-efficient models that keep customer data secure. Koa is post-trained by Salesforce and Nvidia to excel in sales, marketing, and customer-support tasks, offering enterprises an open-weight alternative to closed frontier models like Claude or ChatGPT. Unlike proprietary AI labs that ingest customer data, Koa uses only synthetic data mimicking customer patterns, ensuring no actual customer data leaks. It also reduces token consumption—and thus AI spending—by routing simpler tasks through an AI gateway, reserving frontier models only for complex reasoning. “We’ve built many small task-specific language models… but reasoning has always been something we relied on frontier model providers for. Until now,” said Jayesh Govindarajan, EVP of Salesforce AI. Before Koa, multi-step reasoning prompts were sent to Claude or ChatGPT via Agentforce’s AI gateway. The choice of Nemotron was driven by its sovereign American provenance, state-of-the-art performance, and clear data lineage—contrasting with Chinese open-weight models like Alibaba’s Qwen. Nvidia’s VP of Generative AI Software, Kari Ann Briski, highlighted Nemotron’s “token efficient” architecture for enterprise inference. Salesforce is not abandoning Anthropic or OpenAI; it also announced Claudeforce, a partnership allowing companies to use Claude as their AI interface while data remains in Salesforce’s secure system. Koa will be offered as an alternative within Agentforce, Salesforce’s platform for building customer-service and scheduling agents.