Salesforce Unveils Koa, Its First Reasoning AI Model, With Nvidia

Salesforce and Nvidia introduced Koa at Dreamforce, an open-weight reasoning model post-trained for CRM work that challenges closed frontier AI models.

At its Dreamforce conference, Salesforce announced Koa, the company's first in-house reasoning model, built on Nvidia's open-weight Nemotron model family and post-trained by both companies to handle sales, marketing and customer support tasks. Koa is available as an alternative to external models inside Agentforce, where customers build agents for rote work such as answering service questions or scheduling appointments. Salesforce says customers can select Koa as a managed LLM in the Data Cloud catalogue, enable it org-wide as a model provider, or set it as the standard model in Prompt Builder for individual agents, according to Trending Topics. It is available to pilot customers now, with US general availability planned for the winter and an open beta shortly after.

Koa is a striking example of how enterprise AI needs are pulling away from what the big labs are selling. Salesforce frames it as an open-weight alternative to closed frontier models, trained for specific work tasks rather than hard math problems, with token-efficient inference that cuts spending. Crucially, the post-training used no actual customer data — instead the teams simulated a customer service environment with persona professionals, from irate callers to salespeople closing deals, so the model cannot leak customer information. Salesforce AI EVP Jayesh Govindarajan told TechCrunch the company had always wanted to train an enterprise-grade frontier model but lacked a suitable pre-trained base until Nemotron arrived, calling it the first sovereign, state-of-the-art American model with clear data provenance.

The performance picture is mixed but pointed. According to Trending Topics, Salesforce measures Koa against its own CRM Bench of real-world tasks like updating opportunities and routing cases, reporting three times fewer errors than leading models, 11 percent more precision in calling the right action, 2.1 times greater reliability, and 15 percent better context retention across long conversations. On general benchmarks, Trending Topics reports Artificial Analysis places Nemotron 3 Ultra, roughly 550 billion parameters with 55 billion active, at 23 points — third among US open-weight models but well behind Claude Fable 5.1 and GPT-6 Astra at 53. Nvidia's Kari Ann Briski told TechCrunch its inference architecture is token efficient, citing a 'trifecta' of sovereign AI, time to first token and efficient reasoning.

The move signals a shift beyond Salesforce: open weights let software companies run tailored, auditable models without doing their own pre-training, reducing dependence on frontier-lab price lists. Trending Topics notes Palantir is deploying Nemotron in air-gapped environments for US government agencies through its Sovereign AI Operating System, showing the same appeal of keeping weights and operational knowledge in-house. Salesforce isn't abandoning the labs, though — it announced a partnership with Anthropic called ClaudeForce letting companies use Claude as their AI interface while data stays in Salesforce's system of records, and its AI gateway will route routine tasks to the cheaper in-house model and harder cases to bought-in frontier models.