Salesforce and NVIDIA Introduce Koa, a Reasoning Model for Agentforce
Salesforce developed Koa with NVIDIA for multi-step CRM tasks in Agentforce. The company post-trained it on synthetic scenarios and says it did not use customer data.
Salesforce and NVIDIA Koa is a new specialized reasoning model for the Agentforce platform, introduced by the companies on September 15 at the Dreamforce conference in San Francisco. According to Salesforce, it is the company’s first CRM reasoning model. It was developed in collaboration with NVIDIA by post-training the NVIDIA Nemotron 3 Super model on synthetic scenarios designed for the CRM environment.
The model is already deployed internally at Salesforce, and selected customers are testing it in pilot operations. General availability for U.S. regions is planned for winter 2026, but the company has not announced an exact date or the scope of availability outside pilot programs.
Salesforce and NVIDIA: Koa for CRM Workflows
Koa is designed to perform multi-step tasks in CRM and use tools within the Agentforce environment. Salesforce is expanding an approach in which enterprise workflows rely not only on general-purpose external language models, but also on a model tailored to the specific CRM domain.
Koa is based on NVIDIA Nemotron 3 Super. Salesforce post-trained it on proprietary synthetic scenarios that, according to the company, draw on insights gained during nearly 30 years of CRM deployments. According to Salesforce, customer data was not used in training.
Another stated feature is the model’s operational boundary. According to Salesforce, Koa’s weights, post-training and inference are to remain within its own infrastructure. The announcement therefore also targets enterprise customers addressing data control and model deployment practices when using AI systems.
In-House Benchmark Without Independent Verification
At the launch, Salesforce said Koa achieves comparable or better performance in its CRM Bench benchmark with three times fewer errors. However, this result comes from the company’s own benchmark and has not been independently verified. Another point to watch will be whether Salesforce publishes reproducible results that allow direct comparison with competing models.
Jensen Huang appeared with Marc Benioff at Dreamforce on September 15, and NVIDIA also confirmed the collaboration in its own communications. For NVIDIA, Koa represents an enterprise deployment of the Nemotron model family, while Salesforce gains a model adapted for tasks it wants to address through Agentforce.
What Comes Next for Salesforce and NVIDIA Koa
The next verifiable points will be customer pilot results, the exact general availability schedule and any expansion beyond the United States. It will also be important to see whether Salesforce publishes a more detailed and reproducible evaluation of CRM Bench, including comparisons with other models.
The claim about the composition of the training corpus without customer data will also require independent verification. The announced benefits for agent security, costs or work quality cannot yet be assessed without broader deployment and independent testing.
Sources
- Salesforce – Confirms the announcement of Koa, its basis in Nemotron 3 Super, the synthetic training corpus, operating mode, pilots and planned availability.
- NVIDIA Blog – Confirms Huang’s appearance at Dreamforce, the collaboration with Salesforce and NVIDIA’s description of Koa.
- NVIDIA Research – Documents the technical properties and agentic focus of the Nemotron 3 Super base model.
- TechCrunch – Independently confirms that Koa is among the main Dreamforce announcements and is specialized for enterprise CRM tasks.
Verified and updated: 09/16/2026 06:25
