Salesforce & Nvidia Launch Koa — Enterprise's First Open-Weight Reasoning Model

Salesforce and Nvidia unveiled Koa at Dreamforce 2026 on September 15, 2026 — an open-weight reasoning model built specifically for enterprise sales, marketing, and customer support. Unlike Claude or ChatGPT, Koa doesn’t require companies to pipe their proprietary customer data upstream to frontier AI labs. Instead, it runs inside Salesforce’s own infrastructure, trained on synthetic conversations rather than real customer interactions.

The announcement signals a meaningful shift in enterprise AI: rather than treating frontier models as the only option for reasoning-heavy tasks, companies now have a sovereign, token-efficient alternative that keeps data contained — and costs lower. For the broader AI industry, Koa represents the growing appeal of open-weight models as a counterweight to closed systems like Claude and GPT.

What is Koa?

Koa is Salesforce’s first proprietary reasoning model. It is built on top of Nvidia’s Nemotron — an open-weight foundation model — and then post-trained specifically for tasks in sales, marketing, and customer support. Rather than using real customer data for training (which would raise privacy and compliance concerns), Salesforce and Nvidia generated synthetic training data by simulating customer service environments.

“We actually simulated a customer service environment with a persona customer service professional … all the way to a sales professional trying to close a deal,” said Jayesh Govindarajan, EVP of Salesforce AI, during the Dreamforce keynote. The synthetic approach means Koa learns patterns of enterprise dialogue without ever ingesting actual customer records.

Why Koa matters: the end of upstream data leakage

Before Koa, Salesforce’s Agentforce platform routed multi-step, reasoning-heavy tasks through its AI gateway directly to frontier models — Claude, ChatGPT, and others. Every time an agent asked a complex question, the customer’s data left Salesforce’s infrastructure and entered Anthropic’s or OpenAI’s systems. That creates three problems:

• Data privacy risk: Proprietary customer information is uploaded to third-party AI providers
• Higher costs: Frontier models consume more tokens for equivalent tasks
• Lack of provenance: Enterprises cannot verify what training data those models were built on

Govindarajan put it bluntly: “We have no idea what Qwen trains on.” That data-provenance concern applied not just to Chinese models like Alibaba’s Qwen, but to all frontier systems — including those based in the US. The lack of a sovereign, auditable, open-weight base model was the key blocker.

“Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance,” Govindarajan said.

How Koa works with Agentforce

Within Salesforce’s Agentforce platform, customers build AI agents that handle tasks like answering customer-service questions, scheduling appointments, and qualifying leads. Previously, when an agent needed to perform multi-step reasoning, the request was forwarded through an AI gateway to Claude or ChatGPT. With Koa, those same reasoning tasks can now be handled natively inside Agentforce.

Nvidia’s Kari Ann Briski, VP of accelerated computing, described Nemotron (and by extension Koa) as offering “the trifecta of things that you need to have: sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all.” In practice, that means faster responses, lower token bills, and data that never leaves Salesforce’s security perimeter.

The Claudeforce partnership

Importantly, Salesforce has not abandoned Anthropic or OpenAI. Alongside Koa, the company simultaneously announced Claudeforce, a partnership with Anthropic that allows enterprises to use Claude as the AI interface while all customer data remains secured within Salesforce’s own infrastructure. It’s a complementary approach: Koa handles reasoning-heavy tasks internally, while Claudeforce gives companies a Claude-powered experience with enterprise-grade data protection.

This dual strategy reflects a broader industry trend. Companies are looking for ways to harness frontier AI capabilities without ceding control over their most valuable asset — customer data. Open-weight models like Nemotron provide the technical foundation; partnerships like Claudeforce provide the user-facing experience.

Why AI labs should take notice

Koa’s launch carries implications beyond Salesforce. For the first time, a major enterprise customer can route complex AI reasoning tasks to a model that doesn’t require uploading proprietary data. If other companies follow suit — building task-specific reasoning models on open-weight foundations — the frontier labs’ stranglehold on enterprise AI could weaken.

The token-efficiency angle matters too. Koa is described as burning fewer tokens than equivalent Claude or ChatGPT routing, which translates directly to lower costs at scale. For enterprises running thousands of agents across customer support, sales, and marketing, even marginal per-token savings compound quickly.

TechCrunch’s coverage framed the story directly: “Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear.” The fear isn’t that Koa will outperform Claude or GPT on benchmarks — it’s that it makes the current business model of enterprise AI less necessary.

Pricing and availability

Specific pricing for Koa was not disclosed at the Dreamforce announcement. The model is available through Salesforce’s Agentforce platform and is positioned as a cost-effective alternative to frontier-model routing. Salesforce has not yet detailed a public standalone offering — access appears tied to existing Agentforce subscriptions.

Frequently asked questions

Related articles

Comments