Salesforce Built Koa on Open Weights. Here's What That Does to Your AI Margins

Quick answer: On September 15, 2026, Salesforce unveiled Koa at Dreamforce, its first in-house reasoning model, built by post-training NVIDIA's open-weight Nemotron-3-Super-120B with reinforcement learning on synthetic CRM workflows. No customer data was used. Koa now sits inside Agentforce as a cheaper alternative to Claude or ChatGPT, selected automatically by Salesforce's AI gateway. Salesforce hasn't published a head-to-head against frontier models, so "cheaper" is verified and "as good" is still a claim. For anyone shipping an AI feature, the real story is that gross margin on AI features just became a routing decision instead of a price you negotiate once with one vendor.

Quick disclosure, since you are reading this on a directory's blog: this is SaaSCity, a gamified startup directory, and yes, there's a plug for it near the end. The rest of this is the straight version of what Koa actually is, what it proves, and what it doesn't, because the "cheaper reasoning model" headline is doing a lot of work it hasn't earned yet.

The giant that stopped renting its reasoning

Here's the sentence that should stop you mid-scroll if you sell software with an AI feature bolted on: the world's biggest CRM vendor just decided it no longer wants to pay a frontier lab every time one of its agents has to think for more than one step.

That's what happened at Dreamforce in San Francisco on September 15, 2026. Salesforce announced Koa, one of the biggest reveals of the conference, and it is not a new foundation model built in a Salesforce lab. It's a post-trained version of someone else's open-weight base, tuned with reinforcement learning to call the right tools in the right order on CRM-shaped tasks, using fewer tokens than shipping the same request off to Claude or ChatGPT.

Jayesh Govindarajan, EVP of Salesforce AI, put the old constraint plainly: "We've built many small task-specific language models... But reasoning has always been something that we've relied on the frontier model providers for. Until now." The actual blocker wasn't compute or talent. It was provenance. "We have no idea what Qwen trains on," he said, explaining why Salesforce waited for a US-built, state-of-the-art open-weight base with documented data lineage before building on top of one. That base arrived from NVIDIA.

What Koa actually is, in a table

Skip the vibes and look at the numbers. NVIDIA's own Nemotron 3 Super model page, where Nemotron-3-Super-120B was published March 10-11, 2026, is where Koa's foundation comes from, and the specs explain why Salesforce picked it over training from zero.

Spec	Nemotron-3-Super-120B (the base)	What Koa adds
Total parameters	120B	Same architecture, weights updated via RL
Active parameters per token	12B (sparse MoE, hybrid Mamba-Transformer)	Unchanged, still cheap per forward pass
Context window	1M tokens	Inherited
Precision	Pretrained in NVFP4	Inherited
Decoding	Multi-Token Prediction layers	Inherited
License	NVIDIA Nemotron Open Model License, weights and training data published	Salesforce controls the post-trained weights and hosts them on its own infrastructure
Training data for the base	NVIDIA's published pretraining corpus	Public + synthetic CRM workflow data only, no customer data
Throughput claim (vs. GPT-OSS-120B, 8k-in/64k-out)	Up to 2.2x higher, per NVIDIA	Same base, Salesforce didn't change the architecture
RULER score at 1M context	91.75, per NVIDIA	Inherited

Nothing in that table is Salesforce's own invention except the last two rows, and that's the point. The expensive, hard-to-replicate part, a 120B-parameter model with 1M-token context and a mixture-of-experts scheme that activates only 12B parameters per token, was already sitting on NVIDIA's model page for anyone to download since March 2026. Salesforce's contribution was weeks of targeted reinforcement learning, not a pretraining run. For the backstory on how NVIDIA ended up owning this much of the stack, see our piece on NVIDIA acquiring Hugging Face.

How they trained it without touching a single customer record

The training method is the part worth stealing, conceptually, even if you'll never run it at Salesforce's scale.

Salesforce used reinforcement learning with Group Relative Policy Optimization, GRPO, built around what the team calls a "simulation-to-reward" pipeline. The starting point isn't a scraped dataset. It's Agent Script, Salesforce's declarative language for describing what an Agentforce agent should do. Engineers write workflow specs in Agent Script, then expand each into persona-conditioned, multi-turn tasks: an irate customer calling support, a sales rep closing a deal before quarter-end, a claims adjuster working a healthcare case. Scenarios span manufacturing, healthcare, travel and financial services, and each pairs a persona with a mapped sequence of tool calls the agent needs to complete it. The model gets rewarded for successful tool use, not for sounding right.

The dataset's realism comes from 27 years of CRM deployment patterns Salesforce has watched play out across its customer base, distilled into synthetic scenarios rather than lifted from actual tickets. That distinction, patterns learned from experience versus raw customer data in a training set, is the whole legal and sales argument here, and it's worth taking seriously rather than dismissing as spin.

The academic writeup, "Salesforce Koa: An Enterprise Language Model for Agentic Tool Use" (arXiv:2609.15066), submitted September 14, 2026 by Zixiang Chen and the Salesforce Agentforce & AI Research team, reports Koa beating the raw Nemotron base on multi-turn tool use and outperforming unnamed proprietary baselines on agentic reasoning and CRM tasks. That's real, and notably not a comparison against Claude or GPT-5-class models by name. TechCrunch's September 15, 2026 report and ITPro's writeup the same day flag the same gap: independent press has the announcement, not an independent benchmark against the frontier labs Koa is meant to displace.

The gateway is now your margin

Before Koa existed, Agentforce had one move for anything requiring multi-step reasoning: hand it to a frontier model, Claude or ChatGPT, through Salesforce's AI gateway, the routing layer that decides which model handles which request. That gateway is the actual product here, more than Koa itself.

Now the gateway has a choice. High-volume, well-scoped agentic work, resolving support tickets or qualifying leads thousands of times a day, can go to Koa, which Salesforce says burns fewer tokens per task than a frontier model. Anything needing genuine judgment or taste still routes to Claude or GPT. Rohan Kumar, Salesforce's president and chief platform and engineering officer, called Koa a "game changer" for complex multi-step agent work. NVIDIA's Kari Ann Briski framed the value in words that map onto a founder's spreadsheet: "sovereign AI, time to first token, efficient reasoning, for the tokenomics of it all."

That last word, tokenomics, is the one to sit with. If you're paying a frontier lab per token for every request regardless of how trivial it is, your gross margin is whatever that vendor decides to charge this quarter. A router that sends most volume to a cheaper, purpose-tuned model and reserves the expensive one for the requests that actually need it turns margin into something you design. We covered the token-cost math behind agentic workflows in our piece on tokenomics for agentic software engineering, and the same logic applies here. It's also part of why OpenAI is building its own inference chips: whoever controls the cost per token controls the margin on top of it. If you're still deciding which model to standardize on for a given feature, our guide to model fatigue walks through that decision.

This isn't happening in isolation. The same Dreamforce week, Salesforce announced ClaudeForce, a partnership letting companies use Claude as their interface while data stays inside Salesforce's system of record, plus expanded Agentforce access to Amazon Bedrock's model library (Anthropic, OpenAI, NVIDIA) and deeper AWS and Google Cloud integration. Salesforce isn't betting the farm on Koa replacing frontier models. It's betting on a router smart enough to use all of them, and owning one option itself.

"We never trained on your data" is becoming a sales line

Data provenance used to be a footnote in a vendor security questionnaire. It's turning into a headline claim.

Salesforce controls Koa's weights and runs the model on its own infrastructure. In a sales conversation, that lets a rep say the model never trained on customer data and the workload never leaves Salesforce's perimeter, no caveat about a third-party API call to a frontier lab's servers. That's the pitch behind Govindarajan's "we have no idea what Qwen trains on" line: an open-weight base with published training data (Nemotron's license requires it) lets a vendor make a provenance claim it literally cannot make about a closed model or a foreign-trained open model with undocumented sources.

This is the same current pulling European and increasingly US enterprise buyers toward sovereign AI, models whose training data, hosting location and ownership chain are auditable. If you sell into regulated industries, expect "where did the training data come from" to become an RFP checkbox within a year or two. Mistral has been building its entire European pitch around this dynamic; we covered it in Mistral's sovereign open-weight strategy.

What this does not prove

Worth being honest about the gap between the announcement and the evidence.

Salesforce has not published a benchmark putting Koa against Claude or ChatGPT on the same agentic tasks. The arXiv paper compares Koa to the raw Nemotron base it was built from, which it should beat, and to unnamed proprietary baselines. That's a real improvement over the base model, not proof Koa matches frontier reasoning on anything genuinely hard.

Synthetic training data has edges by construction. Scenarios modeled on 27 years of deployment patterns capture common shapes, not your actual customers' weirdest tickets. And because Koa sits behind a gateway that can quietly fall back to Claude or GPT, "Koa handled this ticket" might really mean "Koa was tried first, then escalated." That's sensible engineering, but it means adoption numbers alone won't tell you how often Koa actually finishes the job.

Cheaper tokens on a worse-run task aren't cheaper. If Koa needs two retries and a human escalation to do what Claude does in one pass, the savings evaporate. Nobody outside Salesforce has that failure-rate number yet.

The playbook you can run this quarter

None of this requires being Salesforce. Koa is already running internally (a Slack agent that helps employees find information) and in pilots with Formula 1, Xero, Baxter Credit Union and UChicago Medicine, but the pattern scales down fine.

1. Pick one high-volume, well-scoped workflow. Not "customer support" broadly, a specific task: refund status lookups, lead qualification against a fixed rubric, ticket triage into three categories. Narrowness beats complexity.
2. Log real tool-call trajectories, not prose descriptions. Write down the sequence of API calls and decisions a correct run takes. A spec in plain markdown works fine, no Agent Script required.
3. Start from an open-weight base sized to the job. Nemotron-3-Super-120B is enterprise-scale; smaller open-weight bases exist if you don't need 1M-token context.
4. Try eval-driven prompt optimization before full RL. A tight eval set scored against your logged trajectories, with iterative prompt tuning, gets most of the signal before you touch a training loop.
5. Put a router in front of everything. Even a single if-statement. The goal is that swapping which model handles a task is a config change, not a redeploy.
6. Publish your provenance story. If your model was tuned on your own logs and a documented open-weight base rather than an opaque API, say so on your pricing page. Enterprise buyers are starting to ask.

Who should actually reassess their stack this week

An AI support agent. Paying per-token for a frontier model to answer "where's my order" and "how do I cancel" is the textbook case for a cheaper, narrowly post-trained model behind a router, with the frontier model reserved for genuinely ambiguous tickets.

An AI sales assistant. Lead qualification against a fixed rubric, meeting scheduling, CRM data entry: all high-volume, low-judgment. Deciding which risky deal needs a human touch stays with the expensive model. Close to exactly the split Salesforce built Koa for.

An AI content tool. Less clear-cut. Writing quality is closer to taste than tool-calling accuracy, and taste is where frontier models still earn their token price. Don't downgrade a writing feature just because the routing trend is fashionable; the calculation differs when the output is prose a human reads, not an action a system executes.

The part nobody can post-train away

If a frontier lab can ship a better version of your AI feature next quarter, and an enterprise can now post-train your exact capability onto open weights in a few weeks for a fraction of the cost, the model itself stops being your moat. It probably already wasn't, a lesson Cognition's own $48B valuation debate made explicit for AI-native startups.

What doesn't commoditize as fast is being the thing someone finds at the moment they're comparing options, whether that someone is a human scrolling a directory or a model deciding what to cite when a user asks "what's the best tool for X." That's a distribution problem, not a modeling problem, and the same SEO logic that makes directory listings compound applies to AI citations too. It's the actual reason we built SaaSCity the way we did: a free listing page with a building on a live city map, human editorial review, and a dofollow backlink plus a Monday launch slot (capped at 15 free launches a week) once you add the badge to your site. Quick Pass at $19.99 skips the queue and goes live within 24 hours. Premium at $39.99 adds a written launch post with three dofollow links. Max at $149 adds a homepage pin, billboard and ticker for 30 days if you want visibility while a news cycle is hot. The domain sits at DR 59 at the last Ahrefs refresh, and a free MCP server lets an agent submit for you.

Salesforce just spent real engineering effort proving that owning your model isn't the hard part anymore. Being the answer someone gets when they ask "who does this well" still is.