Salesforce and Nvidia launch Koa, an in-house reasoning model that challenges the big AI labs

At its Dreamforce conference, Salesforce unveiled Koa, its first in-house reasoning model, built on Nemotron, Nvidia's open-weights model. The two companies partnered to post-train Koa specifically on sales, marketing and customer service tasks, a move the company presents as a direct alternative to relying on the big AI labs (the so-called 'frontier labs') for the reasoning capabilities inside its agent platform, Agentforce.

The article frames the launch as a symptom of an underlying divergence between what enterprises need and what proprietary AI labs offer. As described, those labs would rather have companies upload files, code, prompts and feedback directly into their models and agents, spending large sums of money in the process. Salesforce, by contrast, offers its enterprise customers an alternative with several features: an open-weights model rather than closed ones; trained for concrete work tasks instead of for solving extremely hard math problems; one that has not ingested real customer data and therefore cannot leak it to third parties; one that cuts AI spending by using fewer tokens to do the same job; one that can be routed automatically through an AI gateway as needed; and one that respects each customer's data requirements and security, built into Salesforce's own infrastructure.

Koa will be offered as an alternative to the other models available in Agentforce, the platform where Salesforce customers build agents for routine tasks such as answering customer service questions or scheduling appointments. Jayesh Govindarajan, executive vice president of AI at Salesforce, explained that the company had already developed numerous small, task-specific language models, but that reasoning had always been an area where it depended on frontier model providers such as Claude or ChatGPT, to which Agentforce routed queries that required solving long or multi-step tasks. With Koa, according to Govindarajan, that is no longer necessary — 'until now'.

According to the same executive, the reason Salesforce had not trained its own enterprise-grade frontier model earlier was not a lack of will but the absence of a suitable pre-trained base model: until Nemotron came along, there was, in his words, no US-origin 'sovereign' pre-trained model that was both state of the art and had clear data provenance. Govindarajan contrasted this with Qwen, the popular Chinese open-weights model developed by Alibaba, about which he said: 'We have no idea what Qwen is trained on'.

Koa's post-training process — turning a general-purpose model into an expert in sales and customer service — was carried out, as explained, without using real Salesforce customer data. Instead, the company and Nvidia created synthetic data that mimicked those customers' patterns. Govindarajan described how they simulated a complete customer service environment, with personas ranging from a support professional facing angry customers calling the contact center to a sales professional trying to close a deal.

On Nvidia's side, Kari Ann Briski, vice president of enterprise generative AI software, stressed that Nemotron provides an inference architecture designed to be token-efficient, which she described as the necessary combination of three elements: sovereign AI, time to first token and efficient reasoning, key factors in what she called the system's 'tokenomics'. The underlying idea is that Koa should be cheaper in tokens consumed — and better suited to customers' real tasks — than sending those same queries to Claude or ChatGPT.

The article clarifies, however, that Salesforce is not abandoning Anthropic or OpenAI: in parallel it announced a partnership with Anthropic, dubbed Claudeforce, that lets companies use Claude as an AI interface while their data stays within Salesforce's system of record, protected by its own infrastructure.

Taken together, the Koa case illustrates a broader trend: major enterprise software vendors are seeking their own specialized, more token-efficient reasoning models, built on open foundations such as Nemotron, rather than relying exclusively on the AI labs' closed models for every task that requires reasoning. That could pressure those labs to compete not only on general capability but also on cost per task and on data governance guarantees, against specialized alternatives built in-house by enterprise vendors themselves.