Models & agents

Salesforce releases Koa, a reasoning model built on Nvidia's open-weight Nemotron, for Agentforce

September 16, 2026 at 7:30 AM PT

Salesforce Koa announcement graphic

Image: Salesforce

Why it mattersAn enterprise buyer now has a task-specific reasoning model that stays inside its own vendor's trust boundary, cutting one reason to route work to Claude or ChatGPT.

Salesforce introduced Koa at Dreamforce on 2026-09-15, a family of reasoning models built by post-training Nvidia's open-weight Nemotron on a synthetic dataset that mimics customer relationship management work. Koa is Salesforce's first reasoning model, and it becomes an option inside the Agentforce platform that customers already use to build agents.

Salesforce says Koa is trained to reason through the long-running, multi-step tasks Agentforce customers want an agent to do, such as generating leads, resolving service cases, and scheduling follow-ups. Before Koa, Agentforce routed those prompts to a frontier model like Claude or ChatGPT through its AI gateway, per Salesforce EVP of AI Jayesh Govindarajan speaking to TechCrunch. He told TechCrunch that reasoning was the one capability Salesforce had relied on frontier model providers for, "until now."

How they trained it, and what it does not contain

Salesforce and Nvidia did not use any real customer data. They built the training corpus from synthetic scenarios that simulate the reasoning, tool use and decisions Agentforce agents perform across the customer lifecycle, spanning 14 or more industries, according to the Koa page on salesforce.com. Each scenario pairs a persona with tasks and maps the exact sequence of actions to complete them.

Because Nemotron is open weight, Salesforce controls the weights and runs Koa entirely inside its own infrastructure, the vendor page says. Salesforce says no customer data crosses that trust boundary during training or inference.

Where it plugs in

Koa is available three ways inside Agentforce, per the vendor page: as an option in the Data Cloud generative models catalogue, as an org-wide model provider in Setup, and per agent or sub-agent in Agentforce Builder. It becomes the fourth model provider available in Setup, alongside the existing options, and customers opt in.

Salesforce says it is already using Koa internally on an Employee Agent in Slack, and it is moving into customer pilots with 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.

Two claims to attribute directly to the vendors and read as marketing rather than measurement. Salesforce says Koa is meant to be better at CRM tasks than sending them to Claude or ChatGPT, and cheaper in tokens burned. Nvidia's Kari Ann Briski, VP of Generative AI Software for Enterprise, told TechCrunch that Nemotron's architecture is built for token-efficient inference. Neither company published a benchmark number, only the claim.

Alongside Koa, Salesforce announced Claudeforce, a separate partnership with Anthropic that lets companies use Claude as their interface while data stays in Salesforce records. So Koa is an addition to the model choices inside Agentforce, not a replacement for the frontier models already routed there.

Source

Vendor page: Salesforce Koa. Reporting: TechCrunch, 2026-09-15.

Source: Salesforce

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

More from AI News

Nvidia says a 30B Nemotron fine-tuned on its supply-chain data scored 86.7 percent, versus 55.5 for the 550B Nemotron

Nvidia and Palantir fine-tuned a 30B Nemotron 3.5 Lightning on Nvidia's own supply-chain decisions and, according to Nvidia, it scored 86.7 percent on the internal allocation task versus 55.5 for the 550B Nemotron 3 Ultra, a model 18 times its size.

Source: PressModels & agents

The Information reports Salesforce is letting customers pay for AI based on revenue it grows or costs it cuts

The Information reports Salesforce is starting to let businesses choose how they pay for Agentforce, including custom contracts priced on deals closed or service interactions automated.

Source: PressGo-to-market

Meng Zhang's reasoning-prefill rerun shifts Qwen 3.8's answers 18 points toward GPT-5.5 Pro

Prefilling Qwen 3.8 A95B with the first one percent of GPT-5.5 Pro's reasoning tokens moved its answers 18 percentage points closer to GPT-5.5 Pro's, on a private set of 45 problems.

Source: Hacker NewsModels & agents