Salesforce launches Koa reasoning model built on Nvidia Nemotron
Salesforce has introduced Koa, its first reasoning model, at its Dreamforce conference. Built on Nvidia’s open-weight Nemotron model, Koa was post-trained by the two companies for sales, marketing and customer-support tasks and will be offered through Salesforce’s Agentforce platform.
Agentforce is Salesforce’s environment for building agents that handle routine work such as answering customer-service questions and scheduling appointments. Until now, when an agent had to handle a long-running or multi-step task requiring reasoning, Agentforce routed the request through its AI gateway to frontier models including Anthropic’s Claude or OpenAI’s ChatGPT.
An enterprise-focused reasoning option
Jayesh Govindarajan, executive vice-president of Salesforce AI, said Salesforce already had smaller task-specific language models in the Agentforce portfolio, but had relied on frontier providers for reasoning. Nemotron supplied a pre-trained starting point that Salesforce viewed as state of the art, available in the US and accompanied by clear data provenance.
Koa is positioned as an alternative within the models available to Agentforce customers, rather than as a replacement for every external model. The AI gateway can select a model for a request depending on the need, allowing reasoning workloads to be directed to Koa or to other available models.
Synthetic training data and token efficiency
Salesforce and Nvidia did not use Salesforce customer data to post-train Koa. Instead, they created synthetic data intended to mimic customer patterns. Govindarajan described simulated customer-service settings with personas ranging from irate callers to sales professionals seeking to close deals.
That approach is intended to provide a model tailored to enterprise work while avoiding the use of actual customer records in its training. Salesforce also says Koa follows customer data requirements and uses security embedded in its platform. Those controls align with customer-led Salesforce AI roadmap risks as customer involvement shapes release priorities and the associated risks.
Nvidia’s Kari Ann Briski said Nemotron has an inference architecture designed for token efficiency. Salesforce therefore presents Koa as a way to perform targeted reasoning tasks with fewer tokens than routing the same work to Claude or ChatGPT, alongside benefits around time to first token and sovereign AI requirements.
Frontier partnerships remain in place
The launch does not end Salesforce’s relationships with frontier-model vendors. Salesforce also announced ClaudeForce, a partnership with Anthropic that lets companies use Claude as their AI interface while keeping their data in Salesforce’s system of record and under its infrastructure.
For businesses using Agentforce, the immediate implication is to match the model route to the task: assess Koa for specialised sales, marketing and support reasoning where synthetic post-training, data handling requirements and token use matter, while retaining other models where they remain appropriate.