Salesforce launches Koa CRM reasoning model built on Nvidia's Nemotron

Salesforce is getting on the Nvidia Nemotron bandwagon with Koa, a CRM reasoning model that was built on 27 years of a proprietary synthetic data set modeled on CRM deployments.

Koa will be used in Salesforce's Agentforce platform. Koa was built on Nvidia Nemotron 3 Super. With Koa, Salesforce is looking to harness enterprise processes, workflows and policies and encapsulate them in a model.

Enterprise software providers are increasingly leveraging Nemotron models to train them on enterprise-specific use cases. Why? These software providers are keeping their AI costs in check and getting better accuracy from open models.

Indeed, Salesforce said Koa was measured in Salesforce's CRM Bench, which benchmarks CRM use cases and tasks, and found Koa matched or exceeded performance compared to leading models with 3x fewer errors. Salesforce controls the model weights and performed post-training inference within its trust boundary.

Palantir was among the first software providers to bet big on Nemotron and it's not a stretch to see enterprises coalesce around Nvidia's Nemotron family and open models broadly.

• Why you should arbitrage your AI models like your vendors do
• Nvidia's supply chain integrates Palantir
• Nvidia’s $12.93 billion Hugging Face acquisition validates open weight models

For Salesforce's part, Koa is already used inside Salesforce to power an Employee Agent. Customer pilots include 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero.

According to Salesforce and Nvidia, Nemotron-based models will also be brought into Missionforce for government customers who need private clouds, air-gapped networks and secure environments. Salesforce said post-trained Nvidia models will be used for agents in Missionforce Operations, which will aim to automate government workflows, in October.

Key points about Koa include:

• No customer data was used to train Koa.
• The training data was built on synthetic scenarios that reflect what Agentforce agents would see.
• Use cases include generating leads, qualifying opportunities and resolving service cases.
• The training for Koa includes more than 14 industries and each scenario has a persona, sequence of actions and tool calls needed to complete a task.
• Salesforce's post training leveraged Nvidia NeMo RL and NeMo Gym.
• Koa will be available in Agentforce later in 2026 in the US.

For more on open models see:

• Open-Weight Models Are Gaining Ground in Enterprise AI
• Get ready for US open LLMs and just in time
• Rightsizing open models may cut your AI inference spend
• The most important LLM benchmark today: Price