Salesforce Launches First Reasoning Model Koa with NVIDIA
Salesforce launched Koa, its first self-developed reasoning model, at Dreamforce 26, announced jointly by CEO Benioff and NVIDIA CEO Jensen Huang. Positioned for real enterprise workloads, it powers Agentforce and, per research team member @huanwang, is the payoff of a three-year-old bet that teaching models to correctly execute actions (function calling) would be key to AI agents.
Confirmed
- Koa was built with NVIDIA, post-trained on the open-source Nemotron-3-Super-120B, runs within Salesforce's trust boundary, and drives Agentforce.
- Per @omarsar0, Salesforce extended its declarative Agent Script specification files into simulated training environments, doing RL with its own agent configuration files.
- Per @silviocinguetta, Koa's training drew on 27 years of Salesforce enterprise workflow data.
- Per @huanwang, the technical lineage started with the open-sourced xLAM large action model series (models and data), evolved through xLAM-2 and APIGen-MT, and landed this week as Koa in partnership with NVIDIA, moving long-lab-bound technology into production.
Why it matters
- It is a replicable template for large SaaS companies turning their declarative agent definitions directly into RL training data, showing a viable path for enterprise-specific models.
- It also shows how basic research bets like "function calling three years ago" can end up in enterprise products: Koa targets actual 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 ·
- Salesforce turns its own agent config files into RL environments to train enterprise model Koa — omarsar0 ·
- Salesforce launches Koa, an enterprise reasoning model post-trained with NVIDIA on 27 years of workflows — silviocinguetta ·
- [source] Salesforce turns its own agent config files into RL environments to train enterprise model Koa — omarsar0 · 2026-09-16
- [source] 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
- 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