Skip to main content

Salesforce and NVIDIA Unveil Koa: Open-Source AI Model Takes On Closed-Source Giants

Salesforce and NVIDIA Join Forces to Launch Koa, an Open-Source Reasoning Model

At the recent Dreamforce conference, Salesforce pulled back the curtain on Koa, its first-ever reasoning large model. Built on NVIDIA's open-source Nemotron foundation and fine-tuned by both companies, Koa is designed to supercharge sales, marketing, and customer service tasks.

For years, enterprises have faced a tough trade-off: hand over all your internal documents, code, and user feedback to closed-source AI labs, and you get powerful models—but also hefty costs and nagging data security worries. Koa flips that script. It's an open-weight model that never ingested a single byte of real customer data during training. That means no risk of leaking sensitive info.

Image

Cheaper, Safer, and Built for Business

Thanks to NVIDIA's Nemotron architecture, Koa sips far fewer tokens than third-party closed-source models when tackling the same business tasks. Companies can also route requests flexibly through an AI gateway, matching resources to actual demand. And because Koa's security is baked right into the Salesforce platform, it meets the strict data management rules that enterprise customers demand.

Previously, if an agent needed to handle long, multi-step reasoning, Salesforce's Agentforce platform had to forward the request to external heavyweights like Claude or ChatGPT. Jayesh Govindarajan, Salesforce's SVP of AI, admitted the team had long wanted to build a high-performance enterprise reasoning model, but the lack of a U.S.-made base model with top-tier performance and clear data provenance kept getting in the way. Nemotron changed that. Its training used entirely self-generated synthetic business datasets—simulating everything from customer complaints to sales promotions—giving Salesforce the trustworthy foundation it needed.

A Diversified Strategy, Not a Monogamous One

Even with Koa in its pocket, Salesforce isn't putting all its eggs in one basket. The company recently announced ClaudeForce, a project with Anthropic that lets enterprise users interact with Claude as a front-end AI while keeping all core business data safely inside Salesforce's system of record.

So what does Koa's arrival really signal? For one, it shows that open-source alternatives can be both powerful and practical for serious enterprise work. It also puts pressure on closed-source AI giants to justify their hefty price tags and data-hungry practices. And for businesses tired of shipping their sensitive data off to third parties, it offers a compelling middle path: the performance of a cutting-edge reasoning model, without the black-box risks.

As AI continues to weave itself into the fabric of everyday business, moves like this one could shift the balance of power—away from a few closed-source gatekeepers and toward a more open, transparent ecosystem. Whether Koa lives up to its promise will depend on real-world results, but for now, it's a bold step in a direction many enterprises have been quietly hoping for.

Key Points

  • Salesforce launched Koa, its first reasoning large model, built on NVIDIA's Nemotron open-source foundation.
  • Koa was trained without any real customer data, eliminating privacy and compliance risks.
  • It consumes fewer tokens than closed-source alternatives, cutting costs for enterprises.
  • Salesforce still collaborates with Anthropic on ClaudeForce, showing a diversified AI strategy.
  • The move challenges closed-source AI giants by offering a powerful, transparent open-weight option.