Salesforce builds Koa, its first reasoning model, by post-training Nvidia's open-weight Nemotron
Rather than licensing a frontier model, Salesforce fine-tuned an open-weight base on synthetic CRM data and keeps inference in-house. It is a template for how domain-specific models are now made.
Salesforce has announced Koa, which it describes as its first reasoning model, built by post-training Nvidia's open-weight Nemotron 3 Super, according to a report by TechCrunch on 15 September 2026 and a Salesforce press release published the same day. The model is intended for Agentforce, the company's platform for AI agents, and is tuned for sales, marketing and customer-support work.
According to Salesforce, Koa was produced with supervised fine-tuning followed by reinforcement learning using Group Relative Policy Optimisation, using Nvidia's NeMo RL, NeMo Gym and NeMo AutoModel tooling. The training data was a proprietary synthetic dataset that the company says reflects 27 years of CRM experience, with scenarios simulated across more than 14 industries including manufacturing, financial services, healthcare and travel. TechCrunch reported that the synthetic data included personas such as irate customers and sales professionals closing deals. Salesforce said no customer data was used in training.
On its own CRM benchmark, Salesforce says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors. The company controls the model weights, runs inference entirely within its own infrastructure, and says no customer data crosses its trust boundary during inference. TechCrunch reported that the model is routable through Salesforce's AI gateway and uses fewer tokens than frontier models for the same work.
Jayesh Govindarajan, executive vice president of Salesforce AI, told TechCrunch that reasoning had always been something the company relied on frontier model providers for, "until now". Koa is available to select pilot customers, including 1-800Accountant, Baxter Credit Union, Engine, Formula 1, UChicago Medicine and Xero, with general availability expected in winter 2026 in US regions. Salesforce also announced a partnership with Anthropic that lets customers use Claude while keeping data inside Salesforce's systems.
Why it matters
For years the working assumption was that reasoning-capable models could only come from a handful of frontier labs, and everyone else would rent them. Koa is an example of a large software company taking an open-weight base, spending its effort on domain-specific post-training and evaluation, and hosting the result itself. The pitch is not that Koa is the smartest model available; it is that a narrower model with controlled data flows and lower token costs is good enough for a defined job. If that holds, the centre of gravity in enterprise AI shifts from who owns the biggest model to who owns the best task data and evaluation.
For years the working assumption was that reasoning-capable models could only come from a handful of frontier labs, and everyone else would rent them.
Open-weight base
Nvidia Nemotron 3 Super ships its parameters, so Salesforce can continue training and host it.
Synthetic CRM data
Simulated scenarios across 14+ industries and personas; no customer records used.
Supervised fine-tuning
The model learns the format and behaviour of CRM tasks from example input-output pairs.
Reinforcement learning (GRPO)
Sampled answers are scored against each other and the model is nudged toward higher-scoring reasoning.
In-house inference
Salesforce holds the weights and runs the model inside its own infrastructure, behind its AI gateway.
What you can learn from this
- Open weights make post-training possible. An open-weight model ships its trained parameters, so a third party can load them, continue training, and deploy the result on its own hardware. A closed model only offers an API. Nemotron 3 Super gave Salesforce a starting point that already understood language and code, so the company's work went into teaching CRM behaviour rather than teaching English. This is the pattern behind most domain models today: a general base plus specialised post-training.
- Supervised fine-tuning teaches the format; reinforcement learning teaches the judgement. SFT shows the model examples of inputs paired with good outputs and adjusts it to reproduce them, which is efficient for learning structure, tone and tool-call formats. Reinforcement learning then lets the model generate its own answers, scores them, and updates it toward higher-scoring behaviour. GRPO is a variant that compares a group of sampled answers against each other to compute the reward signal, avoiding the separate value model that older methods needed. The combination is how most reasoning models are now trained.
- Synthetic data is a design choice with trade-offs. Salesforce could not train on customer records, so it simulated them. Synthetic data lets you cover rare scenarios deliberately, label them precisely, and avoid privacy problems. The risk is that the simulation encodes the assumptions of whoever built it, and the model learns those rather than reality. The safeguard is evaluation on held-out real tasks, which is why the benchmark claim matters more than the training recipe.
- A benchmark you wrote yourself is evidence, not proof. Three times fewer errors on an internal CRM benchmark is a meaningful signal, but a vendor's own benchmark is tuned to its own product and use cases. When you evaluate any model, ask what the tasks were, who wrote them, whether the compared models were run with comparable prompts and tools, and whether the test set could have leaked into training. Reading benchmark claims sceptically is a core skill, not cynicism.
- Token efficiency is a cost and a latency story. A reasoning model that thinks in fewer tokens answers faster and costs less per task. For an agent that makes many calls per workflow, that compounds quickly. It also explains the sovereign hosting angle: when you run inference yourself, every token is spent on your own GPUs, so efficiency stops being a line on someone else's invoice and becomes your capacity plan.
We teach this
How to use this in practice
- Run an open-weight reasoning model locally. Install Ollama or a similar runner and pull a small open model from the Nemotron, Qwen or Llama families that supports a reasoning mode. Ask it a multi-step question and inspect the intermediate reasoning it produces. Done looks like a working local endpoint and a note on how many tokens it spent thinking versus answering.
- Write a ten-item evaluation for one task you care about. Choose a job such as drafting a support reply or extracting fields from an email, write ten realistic inputs and the answer you would accept for each, and score two different models against them. Done looks like a small table with a pass rate per model and a sentence on where each one failed. This is the discipline behind Salesforce's CRM benchmark, at desk scale.
- Fine-tune something small. Using a free notebook and a parameter-efficient method such as LoRA, fine-tune a small open model on a few hundred examples of a narrow format, for example turning meeting notes into action items in a fixed schema. Compare outputs before and after. Done looks like the tuned model following your format reliably where the base model did not.
- Trace a data flow. Take any AI feature you use in a personal project or at work and draw where the prompt goes, where inference runs, what is logged, and who can read it. Mark the trust boundary. Done looks like a one-page diagram that would let you answer a customer's question about where their data went.
Sources
- Salesforce and Nvidia's new reasoning model is everything the AI labs should fear — TechCrunch
- Announcing Koa: Salesforce's First CRM Reasoning Model, Built on NVIDIA Nemotron — Salesforce
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report