Salesforce · 2026-09-15 · major
Salesforce Koa — a CRM reasoning model built on NVIDIA Nemotron
Salesforce Koa is Salesforce's first reasoning model, post-trained from NVIDIA's open-weight Nemotron 3 Super with reinforcement learning. Koa beats its base model on multi-turn tool use and is in pilot inside Agentforce.
Salesforce Koa is a CRM-focused reasoning model that Salesforce post-trained from NVIDIA's open-weight Nemotron 3 Super.
Quick facts
| Maker | Salesforce, with NVIDIA |
|---|---|
| Base model | NVIDIA Nemotron 3 Super (120B, open-weight) |
| Training | SFT then reinforcement learning (GRPO), no customer data |
| CRM Bench | 0.86 weighted average (base model 0.84) |
| Availability | Agentforce pilot now; general availability expected winter 2026 in U.S. regions |
| Also shipping in | Missionforce Operations, generally available now |
| Price | Not disclosed |
Benchmarks
| GPT-5.5 | 83.99% | |
|---|---|---|
| Claude Opus 4.8 | 74% | |
| Salesforce Koa | 69.41% | |
| Nemotron-3-Super-120B | 68.64% | |
| GPT-4.1 | 54.48% |
| Claude Opus 4.8 | 78.18% | |
|---|---|---|
| GPT-5.5 | 67.63% | |
| Salesforce Koa | 66.63% | |
| Nemotron-3-Super-120B | 64.73% | |
| GPT-4.1 | 53.96% |
What is it?
Koa is the first reasoning model Salesforce has built itself, announced on 15 September 2026 at Dreamforce and aimed at sales, marketing and customer-support work. The starting point is NVIDIA Nemotron 3 Super, a 120-billion-parameter open-weight model. Salesforce then post-trained it on a synthetic dataset drawn from close to 27 years of its own CRM deployments across more than 14 industries, and says no customer data went into training.
How does it work?
Post-training runs supervised fine-tuning followed by reinforcement learning with Group Relative Policy Optimization, using NVIDIA NeMo tooling. The piece Salesforce calls new is a simulation-to-reward pipeline: a workflow specification, written in the same Agent Script language used to build Agentforce agents, is expanded into many persona-conditioned multi-turn conversations, and the reward is whether the agent actually resolved the task with the right tool calls rather than whether its wording looked right.
Why does it matter?
Multi-turn tool use is where enterprise agents break, and that is exactly where Koa's gains over its base model land — 69.41 versus 68.64 on Tau2Bench and 0.86 versus 0.84 on Salesforce's own CRM Bench. The wider point is who is doing the training: an application vendor took an open-weight model and specialised it rather than buying a frontier API. Salesforce Koa still trails GPT-5.5 and Claude Opus 4.8 on the public tables, so this is a specialisation claim, not a frontier one.
Who is it for?
Salesforce admins and enterprise agent builders
Frequently asked questions
- Can I download Salesforce Koa's weights?
- Salesforce announced Koa as a model inside its own products, not as a download. The press release lists Koa for Agentforce pilot customers now, with general availability expected in winter 2026 in U.S. regions, plus Missionforce Operations. Koa's base model, NVIDIA Nemotron 3 Super, is open-weight, but Salesforce has announced no open weights for Koa itself.
- How does Salesforce Koa compare to GPT-5.5 and Claude Opus 4.8?
- Salesforce Koa sits below both on the public tables in its technical report. On Tau2Bench Koa scores 69.41 against 83.99 for GPT-5.5 and 74.00 for Claude Opus 4.8. On CRM Bench the gap narrows sharply: Koa reaches 0.86 weighted average versus 0.90 for GPT-5.5 and 0.87 for Claude Opus 4.8.
- What data was Salesforce Koa trained on?
- Salesforce trained Koa on public and synthetically generated data only, stating that no customer data was used. The synthetic set is derived from close to 27 years of Salesforce CRM deployment knowledge across more than 14 industries, and enterprise task specifications are written in Agent Script, the declarative language used to build Agentforce agents.
- Which companies are testing Salesforce Koa first?
- Salesforce names six pilot customers for Koa: 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero. These organisations get Koa inside Agentforce ahead of the wider rollout, which Salesforce expects to reach general availability in U.S. regions in winter 2026.