A model of its own
Salesforce has built its first reasoning model. Koa, announced with Nvidia on Tuesday at the Dreamforce conference in San Francisco, is designed to let the company’s Agentforce agents work through multistep sales, service and marketing tasks and pick the right tools to finish them.
It is not trained from scratch. Salesforce post-trained Nvidia’s Nemotron 3 Super, an open-weight model that Nvidia describes as having 120 billion parameters, of which 12 billion are active for any given token, in a hybrid Mamba-Transformer mixture-of-experts design.
How it was trained
According to Salesforce, Koa was post-trained on a proprietary synthetic dataset modelled on 27 years of CRM deployments across more than 14 industries, using supervised fine-tuning and reinforcement learning with Group Relative Policy Optimization through Nvidia’s NeMo tools. The company said no customer data was used.
TechCrunch reported that the developers simulated a customer service environment for the training, with personas ranging from irate callers to a sales professional trying to close a deal.
What Salesforce claims
The performance claim is Salesforce’s own. On its internal CRM benchmark, which tests tasks such as updating opportunities and routing cases, the company said Koa “already matches or exceeds leading model performance on CRM actions with three times fewer errors”. It has not published the benchmark or an independent evaluation, and it did not disclose pricing.
The pitch, as TechCrunch framed it, is that a smaller model tuned for one domain can do the work Salesforce customers want from an agent more cheaply than sending the same tasks to a general frontier model.
Availability
Koa is available now to select pilot customers, with general availability planned for winter 2026 in US regions. Salesforce controls the weights and runs inference in its own infrastructure; nothing in the announcement indicates the model will be released. A separate Missionforce Operations product for government and regulated sectors, using post-trained Nvidia models, is due for select customers in October.
“The most valuable thing Salesforce has built isn’t our platform — it’s the accumulated knowledge of how enterprise business actually works,” said Marc Benioff, Salesforce’s chair and chief executive. Nvidia’s Jensen Huang said: “Every company needs useful AI, tailored to its knowledge, expertise, and work.”
Where Anthropic fits
Koa arrives three weeks after Salesforce and Anthropic announced Claudeforce, under which Claude serves as a reasoning model for Agentforce’s Atlas Reasoning Engine and is the default model in several Salesforce products. Salesforce has not said Koa replaces Claude in any of them. The open question for customers is which tasks Salesforce routes to its own model and which it still sends to a frontier lab.
