Salesforce and Nvidia launch Koa, an enterprise reasoning model built on open-weight Nemotron
The AMW Read
First Salesforce reasoning model on Nvidia Nemotron for Agentforce meaningfully updates the enterprise-agent baseline by substituting open-weight, task-tuned reasoning for frontier routing while retaining Claude.
Salesforce and Nvidia launch Koa, an enterprise reasoning model built on open-weight Nemotron
At Dreamforce, Salesforce unveiled Koa, its first reasoning model, developed with Nvidia on Nvidia’s open-weight Nemotron base and post-trained for sales, marketing, and customer-support work. Koa joins Agentforce as an alternative to the frontier models Salesforce has routed through its AI gateway when agents need multi-step reasoning. Post-training used synthetic simulations of service and sales scenarios rather than live customer data, and Salesforce says the model is designed to use fewer tokens than sending the same tasks to Claude or ChatGPT.
The launch marks a sharper split between general-purpose frontier labs and enterprise buyers who want task-specific reasoning, clearer data provenance, and spend control inside their system of record. Salesforce executives cited Nemotron as a sovereign U.S. pre-trained base with state-of-the-art quality and known data lineage—contrasting that with open-weight models such as Alibaba’s Qwen, whose training mix they said is opaque. Nvidia framed the stack around sovereign deployment, time-to-first-token, and token-efficient inference. Salesforce is not exiting frontier partnerships: it also announced Claudeforce with Anthropic so customers can use Claude while data stays inside Salesforce security boundaries.
For builders and investors, the near-term signal is model routing economics inside large enterprise agent platforms. If Koa reliably handles long-running CRM and support workflows at lower token cost, a larger share of Agentforce traffic may stay on Salesforce-controlled open-weight models, while Claude and ChatGPT remain options for harder or customer-preferred cases. Watch whether other system-of-record vendors copy the same pattern—post-train a trusted open-weight base for vertical tasks, keep a gateway to frontier APIs, and sell data residency as the differentiator.
