Salesforce and Nvidia launch Koa, a reasoning model built for business tasks outside frontier labs

Salesforce unveiled Koa this week at its Dreamforce conference—a reasoning model developed jointly with Nvidia that signals a fundamental shift in how enterprises are approaching AI deployment. Rather than funneling work through closed frontier models from OpenAI and Anthropic, Koa offers companies an open-weight alternative optimized specifically for sales, marketing, and customer support workflows.

Why this matters for the AI power balance#

The move highlights a growing divergence between what enterprise customers actually need and what frontier AI labs are designed to deliver. While proprietary model providers have built their business around convincing companies to upload sensitive data, code, and feedback into their systems—at considerable cost—Salesforce is offering a different path: a self-contained model that never touches customer data and runs more efficiently than sending tasks to external APIs.

Jayesh Govindarajan, Salesforce's EVP of AI, explained the rationale to TechCrunch: "We've built many small task-specific language models, but reasoning has always been something we've relied on frontier model providers for. Until now." The shift addresses a core pain point: models trained on synthetic customer service scenarios can handle routine multi-step reasoning tasks without the latency or expense of cloud-based frontier models.

The technical approach#

Salesforce and Nvidia built Koa by post-training Nvidia's open-weight Nemotron model using entirely synthetic data rather than real customer records. The teams simulated realistic customer interactions—from frustrated support calls to sales conversations—to teach the model enterprise-specific reasoning patterns without exposing proprietary customer information.

Nvidia's VP of Generative AI Software, Kari Ann Briski, highlighted Nemotron's efficiency advantage: the architecture optimizes inference to reduce token consumption, making reasoning tasks faster and cheaper than existing alternatives.

The enterprise advantage#

Koa will sit alongside other models in Salesforce's Agentforce platform, where an intelligent "gateway" routes requests to the most appropriate AI tool depending on the task. The reasoning model offers several concrete benefits:

• No data leakage risk — the model never ingests or stores actual customer data