Salesforce and NVIDIA Ship Koa: A Nemotron-Tuned Reasoning Model Built for CRM

The clearest signal in enterprise AI right now isn't a new foundation model — it's who controls the tuning layer sitting directly above one. On September 15, 2026, Salesforce and NVIDIA announced Koa, a CRM-specific reasoning model built by post-training NVIDIA's Nemotron 3 Super on Salesforce data. The result is a co-developed, domain-anchored model designed to run inside Agentforce — and nowhere else.

That last part is the point.

What Koa Actually Is

Koa is not a general-purpose model with a CRM-flavored system prompt. It is a post-trained derivative of Nemotron 3 Super, NVIDIA's base model, refined on Salesforce-specific data to reason over the objects that define enterprise sales and service work: leads, opportunities, service cases. The distinction matters because post-training on domain data changes how a model handles ambiguous, context-dependent decisions — the kind that dominate real CRM workflows, where a lead record is rarely self-explanatory and an open service case carries history that generic reasoning tends to flatten.

Salesforce is positioning Koa as the reasoning engine powering Agentforce, its agentic platform. The model is intended to give Agentforce agents the capacity to triage customer interactions, draft responses, and route work autonomously — not as isolated tasks, but as coordinated steps inside multi-agent pipelines.

The Architecture of the Partnership

The division of labor here is deliberate. NVIDIA supplies the base model and, implicitly, the compute infrastructure underneath it. Salesforce supplies the vertical fine-tuning and the enterprise distribution channel — meaning the installed base of CRM customers who will never build their own model pipelines. Koa is what happens when those two contributions are combined into a single artifact.

This pattern — NVIDIA as upstream model and compute provider, partners as vertical integrators — is increasingly how NVIDIA's software strategy operates at scale. The company doesn't need to win the application layer directly. It needs its models to become the canonical starting point that enterprise partners tune from. Koa is evidence that the strategy is producing real co-developed outputs, not just API partnerships.

For Salesforce, the calculus runs in the opposite direction. Rather than asking enterprise customers to connect Agentforce to an external general LLM — with the procurement complexity, latency, and data-governance friction that entails — Salesforce can now offer a reasoning model that is native to the platform and trained on the same data structures customers already use. That is a meaningful reduction in integration surface area, which matters enormously to the operators and IT leaders who actually sign off on AI deployments.

What Multi-Agent CRM Looks Like in Practice

The framing Salesforce and NVIDIA used in the announcement points toward something more architecturally ambitious than a single smart assistant. They described Koa as a step toward multi-agent workflows inside CRM — where Koa-coordinated agents handle the full arc of a customer interaction: autonomous triage, drafting, and routing, without a human touching each step.

That is a meaningfully different operational model from today's CRM automation, which typically chains discrete rules-based triggers. A reasoning model at the coordination layer can hold context across a workflow, adjust based on intermediate outputs, and make judgment calls that rule engines cannot. Whether Agentforce deployments actually reach that ceiling in practice will depend on how well Koa generalizes across the messy, inconsistently populated data that characterizes real enterprise CRM instances — a problem no announcement resolves.

What the announcement does resolve is the question of where Salesforce intends to compete. Not at the model frontier in the abstract — but at the point where frontier reasoning meets the specific data structures, objects, and workflows that its customers run on every day.

The Bigger Shift

Koa is a concrete example of a trend that will define enterprise AI over the next several years: the commoditization of base models combined with a fierce competition over who owns the fine-tuning and distribution layer closest to the customer's actual data. General-purpose reasoning capability is increasingly available. Reasoning capability that understands your lead scoring logic, your service escalation tiers, and your opportunity stage definitions — and that lives inside the platform your teams already use — is not. That gap is where Salesforce is planting a flag, with NVIDIA's model architecture as the foundation beneath it.