Salesforce Builds Its First AI Model on Nvidia’s Open-Weight Model Nemotron

Set Trending Topics as a preferred source on Google.

Salesforce used its Dreamforce conference to introduce Koa, the company’s first in-house reasoning model. The foundation comes from Nvidia: Koa is built on Nemotron, the chipmaker’s open-weight model family. The two companies post-trained it together so it can handle work in sales, marketing and customer support.

Koa arrives inside Agentforce as an alternative to the models Salesforce has been buying in until now. According to the company’s FAQ, customers can select it as a managed LLM in the Data Cloud generative models catalogue, enable it org-wide in Agentforce setup as a model provider, or set it as the standard model in Prompt Builder for individual agents and sub-agents.

How an Open Base Model Becomes a CRM Model

The technical route is called post-training. Nvidia publishes Nemotron as a pre-trained model along with its weights, the parameters the model learned during training. Anyone holding those weights can keep training the model, run it on their own infrastructure and narrow it to a specific job. That is what Salesforce did, turning a generalist system into one that understands sales conversations, service cases and CRM records.

The training data is the interesting part. Salesforce says the post-training used synthetic scenarios that imitate typical customer patterns, with real customer data left out of the process entirely. “We actually simulated a customer service environment with a persona customer service professional, including irate customers that call into the customer service center, all the way to a sales professional who’s trying to close a deal,” Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch. A model that has never seen real customer data cannot leak it to anyone else, runs the argument.

Why now? Govindarajan points to the missing foundation. For a long time, he says, there was no pre-trained American model that was open, state of the art and transparent about where its training data came from at the same time. That last point ruled out the strongest Chinese open-weight models for Salesforce: “We have no idea what Qwen trains on.”

Where Nemotron Sits in the Rankings

The independent benchmarking platform Artificial Analysis currently places Nemotron 3 Ultra, the largest model in the family with roughly 550 billion parameters and 55 billion active ones, at 23 points on the Artificial Analysis Intelligence Index. Across a field of about 300 listed model variants, that puts it mid-table and well behind the leaders, where Claude Fable 5.1 and GPT-6 Astra share the top spot at 53 points each.

Among open-weight models, Nemotron 3 Ultra ranks 25th out of some 157 listed model families. Chinese labs own the top of that group: GLM-5.3 from Z.ai leads with 45 points, followed by Kimi K3 at 44 and GLM-5.3-Flash at 42, with Qwen models from Alibaba and DeepSeek close behind at 40 points each.

The picture changes inside the American open-weight field. There, Nemotron 3 Ultra sits third, behind the Inkling models from Thinking Machines (26 points) and comfortably ahead of Meta’s Muse Glimmer (18) and Google’s Gemma 4 (17). When the model launched in early summer, Artificial Analysis called it the most intelligent US open-weight model, far ahead of Gemma 4 and gpt-oss-120b at the time and behind Kimi K2.6. The index scores from the two moments are hard to compare directly, because Artificial Analysis keeps raising the difficulty of the benchmark suite.

For Salesforce, a different set of numbers matters: speed and cost. Nemotron 3 Ultra delivers around 204 tokens per second in Artificial Analysis testing, considerably faster than its Chinese rivals. “We have a unique architecture for inference to be token efficient,” said Kari Ann Briski, Nvidia’s VP of Generative AI Software for Enterprise. The mixture-of-experts architecture activates only part of the parameters per request, which cuts the number of reasoning tokens burned in multi-step agent workflows.

The Numbers Salesforce Puts Behind Koa

Salesforce measures Koa against its own CRM Bench, a set of real-world tasks such as updating an opportunity, routing a case or scheduling a follow-up. Koa already matches or exceeds leading model performance on CRM actions with three times fewer errors, according to the accompanying paper. Compared with the general-purpose models used by default today, Salesforce reports 11 percent more precision in calling the right action, 2.1 times greater reliability, and 15 percent better retention of context across long back-and-forth conversations.

Security is the second pillar of the pitch. Koa runs entirely inside the Salesforce trust boundary, sees only the grounding data, records and instructions a customer chooses to share, and is hosted at temperature 0 so responses stay consistent and repeatable. A dedicated serving harness adds a further layer of trust and safety controls around the model itself.

Koa is available to selected pilot customers now, with general availability in US regions announced for the winter and an open beta planned shortly after. In parallel, Salesforce is bringing Nemotron 3 Nano straight into Agentforce, with a context window of one million tokens and the option to run it on premises or in a private cloud for regulated industries.

Palantir Builds on Nemotron, Too

Salesforce is in good company here. Palantir recently introduced an engine that deploys Nemotron models in sealed-off environments, running through its Sovereign AI Operating System of AIP, Ontology, Foundry and Apollo. The target audience is US government agencies and critical infrastructure operators working in air-gapped setups, meaning systems fully isolated from open networks. The appeal of open weights there matches the Salesforce case: an organization that keeps training a model on its own data keeps the weights, and with them the operational knowledge encoded inside.

Two Paths at Once

Koa leaves the door to the frontier labs wide open. Salesforce has just announced a partnership with Anthropic called ClaudeForce, letting companies use Claude as their AI interface while their data stays in the Salesforce system of records. The platform’s AI gateway, which decides what request goes to which model, allows both paths to run side by side: routine and repetitive tasks to the cheaper in-house model, harder cases to a bought-in frontier model.

The move points to a shift that reaches beyond Salesforce. Open weights let a software company run a tailored model without doing its own pre-training, one it can control, audit and operate at a lower cost than a bought-in generalist. Marc Benioff recently described Agentforce as a new dimension of enterprise AI, as Trending Topics reported. With a reasoning model of its own in the portfolio, the platform grows less dependent on the price lists of the big labs.