Salesforce launches Koa, its first reasoning model, built on Nvidia's open-weight Nemotron

Salesforce's first reasoning model

Salesforce has unveiled Koa, its first reasoning model, at its Dreamforce conference. According to TechCrunch, Koa was built with Nvidia on top of Nemotron, Nvidia's open-weight model, and the two companies post-trained it to handle sales, marketing and customer-support work. It will be offered as an alternative to the other models available in Agentforce, the platform where Salesforce customers build agents for routine jobs like fielding support questions or scheduling appointments.

The gap Koa fills is reasoning. Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch that the company already ships many small task-specific language models as part of Agentforce's portfolio, "but reasoning has always been something that we've relied on the frontier model providers for. Until now." Previously, when an agent had to work through a long-running, multi-step task, Agentforce's AI gateway — the routing layer that decides which model handles which request — passed those prompts to a frontier model such as Claude or ChatGPT.

The enterprise pitch

TechCrunch reads Koa as evidence that what enterprises need from AI is drifting away from what frontier labs are selling, since labs would rather have companies upload files, code, prompts and feedback into their hosted models and pay heavily for the privilege. Salesforce's offer to its customers, as reported:

• an open-weight alternative to closed frontier models
• a model trained for concrete work tasks rather than for solving competition-grade math problems
• no real customer data ingested during training, and therefore nothing customer-specific to leak
• lower AI spending, because Koa burns fewer tokens to do the same work
• automatic routing through Salesforce's AI gateway depending on the task
• adherence to each customer's data requirements and security inside Salesforce

Why Nemotron was the starting point

Govindarajan said Salesforce had long wanted to train its own enterprise-grade model but was blocked by the absence of a suitable base. Nemotron changed that: as he put it, it is the first sovereign American pre-trained model that was both state of the art and equipped with clear data provenance. The comparison he drew was with Qwen, Alibaba's widely used open-weight model — "We have no idea what Qwen trains on," he told TechCrunch.

Trained on simulated customers

Post-training took the general-purpose Nemotron base and steered it toward fluency in sales and customer-support work. Notably, no actual Salesforce customer data went into the process: the companies generated synthetic data that imitated customer patterns instead. Govindarajan described simulated service environments staffed by persona agents — including irate callers — as well as sales personas working to close deals.

On efficiency, Kari Ann Briski, Nvidia's VP of generative AI software for enterprise, told TechCrunch that Nemotron brings "a unique architecture for inference to be token efficient," and described sovereign AI, time to first token and efficient reasoning as the trifecta behind the token economics.

Not an exit from frontier labs

Koa does not mean Salesforce is dropping Anthropic or OpenAI. At the same event the company announced ClaudeForce, a partnership with Anthropic that lets companies use Claude as their AI interface while their data remains in Salesforce's system of record, protected by Salesforce's own infrastructure.

Why it matters

Koa is a concrete case of a major enterprise vendor building its own reasoning layer on an open-weight foundation rather than renting that capability from frontier labs. If a task-specific model can match frontier quality on the narrow jobs enterprises actually automate — while using fewer tokens, never training on customer data and offering provable data provenance — the per-token pricing of hosted frontier models comes under real pressure. It also marks open-weight models maturing into enterprise infrastructure: for regulated buyers, where a base model's training data came from is now a purchasing criterion, and Nvidia is marketing Nemotron's American provenance as exactly that. The gateway routing pattern — cheap specialised models first, frontier models only when needed — looks set to become the default architecture for enterprise agents.