Salesforce Unveils Koa, Its First Reasoning Model, Built with NVIDIA to Power Agentforce

At Dreamforce 26, Salesforce launched Koa, the company's first in-house reasoning model, announced jointly by CEO Benioff and NVIDIA CEO Jensen Huang. It is positioned as a model built for real enterprise workflows and designed to power Agentforce.

Confirmed

Koa was built in partnership with NVIDIA, with post-training based on the open-source Nemotron-3-Super-120B. It runs inside Salesforce's trust boundary and powers Agentforce.
According to @omarsar0, the approach is quite instructive: the company extended the declarative Agent Script spec files that define Agentforce agents into simulation-based (training) environments, and used its own agent configuration files for RL.
As relayed by @silviocinguetta, Koa's training leveraged Salesforce's 27 years of enterprise workflow data.
In a retrospective from @huanwang, a core member of the Salesforce research team: nearly three years ago the team made a small bet that "teaching models to correctly execute actions" (function calling) would become a key capability for AI agents. The effort started with the xLAM large action model series, which stayed open source (models and data), evolved through xLAM-2 and APIGen-MT, and this week culminated in Koa through a collaboration with NVIDIA combining Agentforce-related technology—pushing a long-lab-bound technique into production.

Why it matters

It offers a replicable paradigm for a major SaaS company turning its own declarative agent definitions directly into RL training data, showing a viable path for domain-specific enterprise models.
It also shows how a foundational research bet on function calling made three years ago ultimately landed in an enterprise product: Koa focuses on real action execution for CRM agents rather than general reasoning benchmarks.

2026-09-16 ~ 2026-09-18 · 7 related posts

Primary sources

Salesforce unveils Koa, its first reasoning model, with NVIDIA at Dreamforce — huan__wang · 2026-09-18 07:44 (UTC+08:00)
Salesforce turns its own agent config files into RL environments to train enterprise model Koa — omarsar0 · 2026-09-16 12:34 (UTC+08:00)
Salesforce launches Koa, its first reasoning model, built on xLAM-2 and APIGen-MT — huan__wang · 2026-09-18 07:41 (UTC+08:00)

[source] Salesforce turns its own agent config files into RL environments to train enterprise model Koa — omarsar0 · 2026-09-16
Salesforce launches Koa, an enterprise reasoning model post-trained with NVIDIA on 27 years of workflows — silviocinguetta · 2026-09-18
Salesforce ships Koa, its first reasoning model, born from a 3-year-old function calling bet — huan__wang · 2026-09-18
Salesforce's 3-Year xLAM Bet on Function Calling Ships as Koa — huan__wang · 2026-09-18
[source] Salesforce launches Koa, its first reasoning model, built on xLAM-2 and APIGen-MT — huan__wang · 2026-09-18
[source] Salesforce unveils Koa, its first reasoning model, with NVIDIA at Dreamforce — huan__wang · 2026-09-18

1 near-duplicate retellings: huan__wang