Koa: Salesforce and Nvidia release a reasoning model trained on thirty years of CRM deployments

Announced at Dreamforce, Koa has been post-trained since Nemotron 3 Super on synthetic data taken from Salesforce's CRM history, with three times fewer errors claimed on business tasks.

Salesforce and Nvidia presented on September 15th, at the opening of Dreamforce, a model named Koa, cut for specific use: chaining the stages of a CRM work. Qualifying a business opportunity, choosing the following action on a customer service file, calling the right tool at the right time—these are sequences where generalist models are capable but irregular. Koa is not trained since scratch: it is a post-training of Nemotron 3 Super, the base provided by Nvidia, on a synthetic data set built from almost three decades of customer deployments.

On its own CRM evaluation panel, Salesforce asserts that Koa equals or exceeds the best business equity models in the market, with three times less errors. The figure comes from the publisher and requires to be confirmed by independent measures, but the logic of the product is clear: a narrow model, fed by proprietary data, rather than a leading-edge model leased to a third-party laboratory.

The other argument is control. Salesforce holds the weight of the model and performs post-training as an inference within its own trusted perimeter, without the data passing through an external supplier. The same system is extended to Missionforce, offering services to regulated administrations and sectors, with private cloud or isolated network deployments. Koa is open to a small group of pilot customers in Agentforce, for a general availability announced this winter in the United States.