Salesforce used its Dreamforce stage to show off Koa, the first reasoning model it has trained itself, built by post-training Nvidia’s open-weight Nemotron 3 Super.
The two companies announced the model on September 15. Salesforce describes Koa as a CRM reasoning system, aimed at agents that have to work through a long chain of steps and pick the right tool at each one instead of stopping at a single answer.
Until now those requests left the building. When an Agentforce agent needed to reason through a multi-step job, the platform’s gateway routed the prompt to a frontier model such as Claude or ChatGPT. Koa becomes an in-house alternative for the same work.
Jayesh Govindarajan, who runs Salesforce AI, said the company had wanted its own frontier-grade model for years and kept hitting one wall, the absence of a pre-trained base that was state of the art, American and documented. He pointed at provenance directly, saying nobody knows what Qwen trains on.
Training data came from simulation rather than customers. Salesforce built synthetic scenarios across more than 14 industries, pairing a persona with a task list and mapping the actions and tool calls required. Post-training combined supervised fine-tuning with reinforcement learning using Group Relative Policy Optimization, run through Nvidia’s NeMo RL, NeMo Gym and NeMo AutoModel tooling.
Salesforce holds the weights and runs inference inside its own trust boundary. A research paper describing the work went to arXiv on September 14.
