Salesforce Koa: a model built on Nemotron

Koa is Salesforce's first reasoning model, post-trained from NVIDIA Nemotron 3 Super on synthetic data from nearly three decades of CRM. Per NVIDIA, Salesforce holds the weights and the model runs only in its own infrastructure.

In short

The synthetic data teach the model what CRM work looks like without seeing the real customers.
The open model lets the weights stay with Salesforce rather than with an outside provider.
Customer pilots from October; general availability in winter 2026 in US regions.

Checked on1 October 2026Responsible editorTsvetelin IvanovHow we workMethod · Corrections

The most valuable thing Salesforce has is also the most dangerous to touch: nearly three decades of other people's sales and other people's customers.

Train a model directly on that and you gain accuracy and lose trust on the same day. So their answer is a roundabout one. And interesting.

The facts: on 15 September 2026, at the Dreamforce conference, Salesforce and NVIDIA announced Koa, Salesforce's first reasoning model, built for CRM actions in Agentforce. Per NVIDIA, the model was made by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset drawn from nearly three decades of CRM deployments, with scenarios across more than 14 industries; NeMo RL, NeMo Gym and NeMo AutoModel were used. Because Nemotron is open, Salesforce controls the weights, the model runs on its systems and no customer data is used either in training or when the model runs. In the words of Rohan Kumar of Salesforce, not a single byte of customer data was used. In Salesforce's own test, CRM Bench, Koa matches or exceeds leading models on CRM actions, with 3 times fewer errors. The model already runs inside Salesforce in an employee agent in Slack, from October it moves into pilots with customers such as Formula 1, UChicago Medicine and Xero, and general availability is expected in winter 2026 in US regions. On stage Jensen Huang said the share of open models went from 30 per cent at the beginning of last year to around 70 per cent.

Synthetic data do a translator's job here. The model learns from synthetic scenarios across more than 14 industries what the work looks like, without seeing the real people behind it. Whether the synthetic copy keeps enough of the chaos of a real CRM, CRM Bench cannot say, because the test is Salesforce's too.

The open model buys Salesforce something rare: the weights stay at home.

Huang's 70 per cent was said from a stage, with no clarity about what is being measured and a share of what exactly. I do not copy it down as a statistic, but as the mood in the room.

And an honest note from the kitchen. We could not open Salesforce's own announcement; their page refused access even through a real browser. Everything above rests on the blog of NVIDIA, which announced together with them. The first outside eyes on Koa will be the customers in the October pilots.

The backstory NVIDIA claims Nemotron 3 Ultra runs 10 times cheaper than closed models - our story from 8 July.

The visual is generated code art. No third-party images.

Salesforce built its own model on the open Nemotron and says it used not a single byte of customer data

More on this

NVIDIA: its open model runs 10 times cheaper than closed ones on the same task

LFM2.5: 2.6 billion parameters trained for agent work right on the device

NVIDIA released an open model that makes the robot think on the device itself