Entity
Salesforce Koa
Salesforce Koa is a new enterprise language model built by post-training Nvidia's open-weight Nemotron-3-Super-120B with reinforcement learning (GRPO), designed for agentic tool use without customer data (arXiv, September 15, 2026). It was unveiled around Salesforce's Dreamforce 2026 conference, where CEO Marc Benioff appeared alongside the CEOs of Nvidia, Anthropic, and OpenAI discussing AI progress and responsibility (SiliconANGLE, September 15, 2026).
This page was written and is updated by the paper from the source articles listed below. Importance is how central the topic is to current coverage; confidence is how well-sourced the page is; mentions counts the articles that fed it. How the wiki works →
## Overview
Salesforce Koa is an enterprise language model developed by Salesforce, built by post-training the open-weight Nemotron-3-Super-120B foundation model using reinforcement learning with Group Relative Policy Optimization (GRPO) (arXiv, September 15, 2026). Salesforce publicized the model at its annual Dreamforce conference in San Francisco, where CEO Marc Benioff discussed AI progress alongside the CEOs of Nvidia, Anthropic, and OpenAI (SiliconANGLE, September 15, 2026).
## Key Details
- Trained on public and synthetically generated data, explicitly excluding customer data, aimed at improving tool use and agentic capabilities while preserving general-purpose performance (arXiv, September 15, 2026).
- Its distinctive component is a simulation-to-reward pipeline that expands workflow specifications into persona-conditioned multi-turn tasks with task-specific reward signals (arXiv, September 15, 2026).
- Represents Salesforce's continued push into building proprietary/enterprise-tuned LLMs for agentic tool use, complementing its broader Agentforce and agentic AI product strategy.
- Koa's unveiling coincided with Dreamforce 2026, where Benioff spent significant time on AI progress and responsibility, appearing alongside Nvidia, Anthropic, and OpenAI leadership, underscoring Salesforce's positioning within the broader AI industry conversation (SiliconANGLE, September 15, 2026).
## Strategic Significance
This release signals Salesforce's move toward vertically integrated, enterprise-specific foundation models rather than relying solely on third-party LLM providers, positioning it competitively against Microsoft, Google, and other enterprise AI platform vendors building in-house agentic models. The Dreamforce presence alongside Nvidia, Anthropic, and OpenAI CEOs suggests Salesforce is seeking to elevate its profile as a significant player in frontier AI conversations, not just an enterprise software applier of AI (SiliconANGLE, September 15, 2026).
## Related Entities
Built on Nemotron-3-Super-120B (Nvidia-developed base model). Connects to Salesforce's broader Agentforce and enterprise AI initiatives. Announced in conjunction with Dreamforce 2026, an event also featuring Nvidia, Anthropic, and OpenAI leadership.