Salesforce Koa: an open-weight base, closed at the top

Salesforce announced Koa at Dreamforce on September 15 — a reasoning model built with Nvidia and aimed squarely at CRM work. The interesting part is not that Salesforce trained a model. It is where the model came from and where it ends up: the starting point ships with public weights, and the finished thing runs only inside Salesforce.

What actually happened

The research paper went up the day before the keynote. Per the arXiv preprint, Koa is built by post-training Nemotron-3-Super-120B — Nvidia's open-weight model — with reinforcement learning using Group Relative Policy Optimization. Training used public and synthetic data, no customer data. The pipeline the authors describe is the part worth stealing: they convert written workflow specifications into multi-turn tasks, then grade on whether the tool call actually succeeded. Reward on execution, not on text that looks right.

The paper is honest about where it lands. Koa beats what the authors call a strong proprietary baseline and stays below the strongest frontier models, with the clearest gains on multi-turn tool use. No competitor names, no scores in the abstract.

IT Pro reports that the synthetic data was modeled on 27 years of CRM deployment patterns, that Koa runs on Salesforce's own infrastructure, and that Salesforce controls the weights. Pilots include Formula 1, Xero, Baxter Credit Union and UChicago Medicine. Salesforce president Rohan Kumar called it a game changer for complex multi-step agents. No pricing was published.

Why a closed derivative of an open model matters for your business

Nvidia gives away Nemotron-3-Super-120B. Salesforce took it, spent real money on RL, and the output is a platform feature you rent. That is a legitimate business — and it is also a map. The same base is sitting on Hugging Face for you.

You will not out-train Salesforce on CRM. You do not need to. The transferable idea is the reward signal: build a set of your own workflows — quote to invoice, return to refund, lead to booked job — and score a model on whether the tool call completed, not on whether the answer reads well. That harness is worth more than the model choice, and it survives a vendor swap.

Treat Koa as one more entry in the Agentforce model picker. It is fast and CRM-shaped. It is also a model you cannot run, cannot fine-tune, and cannot take with you.

Sources: arXiv: Salesforce Koa, IT Pro, TechCrunch.