Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
Salesforce unveiled Koa at its Dreamforce conference, describing it as the company’s first reasoning model built on Nvidia’s open-weight Nemotron and post-trained for sales, marketing and customer-support tasks. The partners say Koa was trained using synthetic simulations rather than actual customer data and will be offered as an option inside Salesforce’s Agentforce platform.
Why It Matters
Koa signals a shift in enterprise AI toward task-focused, open-weight models that prioritize data provenance and token efficiency over reliance on closed frontier models. For customers, that could mean more control over data, lower operational AI costs, and models optimized for business workflows.
Key Facts
- Announcement: Koa revealed at Salesforce's Dreamforce conference
- Base model: Built on Nvidia's open-weight Nemotron
- Purpose: Post-trained for sales, marketing and customer-support tasks
- Training data: Post-training used synthetic simulations; no actual Salesforce customer data was used
- Platform: Will be offered as an alternative within Salesforce's Agentforce
Salesforce introduced Koa at its Dreamforce event as the company’s first reasoning model, developed in collaboration with Nvidia and based on the open-weight Nemotron architecture. The firms say they post-trained the model specifically for sales, marketing and customer-support workflows, positioning Koa to handle the multi-step, domain-specific reasoning agents often require.
Salesforce frames Koa as an enterprise-focused alternative to general-purpose frontier models. According to company statements, the model provides an open-weight option that adheres to customer data requirements, is designed to consume fewer tokens for the same tasks, and can be routed automatically through the platform’s AI gateway depending on need. Salesforce also emphasized that Koa was trained without using any real customer records, instead relying on synthetic simulations that mimic common service and sales interactions.
Executives described Nemotron as the enabling base model because it offered a pre-trained foundation with clear data provenance and U.S. sovereignty, attributes they said were previously unavailable in other open-weight options. Salesforce’s EVP of AI explained that before Koa, complex reasoning requests from Agentforce were sent to external frontier providers; the Nemotron base made it feasible to build an in-house reasoning model tailored to enterprise workflows without ingesting customer data.
Nvidia highlighted Nemotron’s inference architecture as a driver of token efficiency and faster time to first token, which Salesforce says helps reduce AI spending for customers. Salesforce is not removing third-party options: it also announced a partnership with Anthropic called Claudeforce that lets companies use Claude as an interface while keeping their records under Salesforce’s security and infrastructure. Koa will be one more model choice inside Agentforce for customers building automated agents for routine business tasks.
Keep Reading
Why an AI Slowdown Could Collapse Under Commercial and US-China Pressure
New Italian unicorn Exein rides the physical AI wave
Spotify finally lets parents exclude kids’ music from Wrapped and recommendations
