McCarthy Building Companies trained Nvidia's Nemotron 3 Nano 4B on its own construction data and reportedly topped larger models on risk classification
In AI, the default assumption has been simple: bigger models win. Palantir just presented evidence that a much smaller one can beat them at a specific job, and do it for less money.
The company reported that a fine-tuned, 4-billion-parameter Nvidia model outperformed larger models on a task built by McCarthy Building Companies, at a lower cost.
A small model with a very specific education
The model in question is Nvidia’s Nemotron 3 Nano 4B. McCarthy, a construction firm, fine-tuned it on more than 100,000 labeled requests-for-information, or RFIs.
For the non-builders in the room, an RFI is the formal question a contractor sends when plans, specs, or site conditions don’t line up. They are the paper trail of everything that could go sideways on a job site.
That makes them a goldmine for spotting trouble early. McCarthy used the trained model to classify potential project risks and predict how those risks might hit schedules and budgets.
According to the findings, the tuned model delivered superior performance on that classification work. It also ran at markedly lower operational cost than its larger counterparts.
McCarthy is also using AI in a separate subcontractor contract-drafting workflow. That workflow is expected to save approximately 30,000 work hours annually.
That is a projection, not a banked result.
The rest of the Sovereignty Bootcamp lineup
The McCarthy results were presented at Palantir’s Sovereignty Bootcamp on October 21, 2026. The event focused on what Palantir calls sovereign AI deployments across multiple sectors.
The idea behind “sovereign” AI is control. Customers fine-tune models on their own data, inside their own environment, and keep ownership of both the data and the resulting model.
Defense contractor L3Harris was one of the featured customers. The company reported that early tests produced cost reductions of 95%.
That figure comes from early testing, so it deserves some patience before anyone treats it as a steady-state number.
Tampa General Hospital also shared results. After fine-tuning the Nemotron model to match its own privacy and risk standards, the hospital saw a 22-fold reduction in AI latency.
It also reported a 7% improvement in classification performance.
The three examples span construction, defense, and healthcare. Those are industries with strict data rules, high stakes for errors, and little appetite for shipping sensitive records to a third-party model.
Background: Palantir and Nvidia’s bet on smaller, owned models
The bootcamp built on partnerships between Palantir and Nvidia launched earlier in the year. The pitch pairs Nvidia’s compact Nemotron models with Palantir’s software for deploying AI on proprietary data.
The broader thesis is that task-specific results come from combining a company’s own data with a smaller, specialized model.
What this means for enterprise AI buyers
The results also highlight where the real advantage sits: in the data. McCarthy’s edge came from more than 100,000 labeled RFIs that no competitor or general-purpose model provider has access to.
There are caveats worth watching. The headline numbers come from customer-reported results, and some, like L3Harris’s 95% figure, are tied to early tests.
McCarthy’s 30,000-hour figure is also an expectation rather than a measured outcome.
