Salesforce & Nvidia Launch Koa, Enterprise's First Open-Weight Reasoning Model
Koa is Salesforce's first in-house reasoning model, announced with Nvidia at Dreamforce on September 15, 2026. It takes Nvidia's open-weight Nemotron model and trains it for one job: running the multi-step CRM tasks behind Agentforce agents. We explain how it was built, what the benchmark numbers really say, and what to check before you rely on it.
⚡ Quick facts
- Announced: September 15, 2026, at Dreamforce: Salesforce's first reasoning model, built with Nvidia
- Base model: Nvidia's open-weight Nemotron-3-Super-120B, post-trained by Salesforce with fine-tuning and reinforcement learning
- Training data: synthetic CRM scenarios across 14+ industries; Salesforce says no customer data was used
- Scores: Tau2Bench 69.41, ahead of GPT-4.1 but well behind GPT-5.5 (83.99), on Salesforce's own report
- Availability: pilots now; general availability inside Agentforce later in 2026, US regions first; Koa itself is not downloadable
What Koa is
Koa is a reasoning model Salesforce built for the multi-step work its Agentforce agents do: qualifying leads, routing service cases, scheduling follow-ups and calling the right CRM tools in the right order. Salesforce did not train it from scratch. According to DataNorth, it started from Nvidia's open-weight Nemotron-3-Super-120B and post-trained it with supervised fine-tuning and reinforcement learning (GRPO), using Nvidia's NeMo RL, NeMo Gym and NeMo AutoModel tools.
The point is control and cost. Salesforce says it holds the weights and runs inference inside its own trust boundary, so Agentforce does not have to send every step of an agent's work to an outside model provider.
How it was trained
Salesforce says Koa was trained on synthetic data, not customer records. As Constellation Research describes it, each scenario has a persona, a sequence of actions and the tool calls needed to finish the task, across more than 14 industries, modelled on what Agentforce agents see in real deployments. That matters for buyers: it avoids the question of whether your CRM data trained someone else's model, but synthetic scenarios can also miss the messy edge cases of a real org.
The benchmark numbers, read carefully
All the scores come from Salesforce's own technical report, summarised by DataNorth:
- Tau2Bench (agent tasks): Koa 69.41, versus 68.64 for the Nemotron base model and 83.99 for GPT-5.5.
- BFCL (function calling): Koa 66.63; the best rival in Salesforce's table, Claude Opus 4.8, scored 78.18.
- CRM Bench: Koa 0.86 against 0.90 for GPT-5.5.
So Koa is a modest step up from its base model and trails the best frontier models on general agent benchmarks. Salesforce's argument, covered by CXOToday and Techstrong.ai, is that a specialised model it controls is good enough for CRM work. DataNorth notes that the rival models' versions and settings were not specified, so treat the comparison as indicative, not controlled. Price, context window and latency figures have not been published.
Who can use it, and when
Koa already runs Salesforce's internal Employee Agent and is in pilot with customers including 1-800Accountant, Baxter Credit Union, Formula 1, UChicago Medicine and Xero. General availability inside Agentforce is expected later in 2026, in US regions first. The “open-weight” label belongs to Nemotron, not Koa: as Trending Topics points out, Salesforce built on an open model, but Koa itself is a managed model inside Agentforce that you cannot download or self-host.
Our take: five questions to ask before switching an agent to Koa
- Does your workload look like the benchmark? Run your own top 20 agent tasks on Koa and on your current model, and count completed tasks and wrong tool calls, not just answer quality.
- What does it cost per conversation? Salesforce has not published pricing; ask for the Agentforce rate with Koa versus a frontier model before assuming savings.
- Where does it run? GA starts in US regions. If your data must stay in another region, check the roadmap first.
- Can you fall back? Keep the option to route hard cases to a stronger model; on Tau2Bench the gap to GPT-5.5 is about 15 points.
- Who owns the failure? Synthetic training reduces data-privacy risk but not error risk. Keep human review on actions that change records or contact customers.
Sources
- DataNorth: Salesforce launches Koa, a CRM reasoning model
- Constellation Research: Salesforce launches Koa CRM reasoning model built on Nvidia's Nemotron
- Techstrong.ai: Salesforce, Nvidia debut Koa reasoning model for CRM
- CXOToday: Salesforce, Nvidia launch new reasoning model Koa
- Trending Topics: Salesforce builds its first AI model on Nvidia's open-weight Nemotron
Frequently asked questions
What is Salesforce Koa?
Koa is Salesforce's first reasoning model, built with Nvidia and announced on September 15, 2026. It is designed for multi-step CRM tasks inside Agentforce, such as qualifying leads, routing cases and calling CRM tools.
What model is Koa based on?
Koa is Nvidia's open-weight Nemotron-3-Super-120B, post-trained by Salesforce with supervised fine-tuning and reinforcement learning using Nvidia's NeMo tools.
Was Koa trained on customer data?
Salesforce says no. It was trained on synthetic CRM scenarios covering more than 14 industries, each with a persona, a sequence of actions and the tool calls needed to complete the task.
How does Koa compare with GPT-5.5 and Claude?
On Salesforce's own report, Koa scored 69.41 on Tau2Bench versus 83.99 for GPT-5.5, and 66.63 on BFCL versus 78.18 for Claude Opus 4.8. It beats its base model slightly but trails the best frontier models.
Is Koa open source or downloadable?
No. The base model, Nemotron-3-Super-120B, is open-weight, but Koa itself is a Salesforce-managed model available only inside Agentforce.
When will Koa be available?
Koa is in pilot with customers such as Formula 1 and Xero. Salesforce expects general availability inside Agentforce later in 2026, starting in US regions. Pricing has not been published.