Palantir Chart Puts Fine-Tuned Nemotron at 73% on RFI Task

Palantir’s Monday, Nov. 2 business update put a McCarthy Building Companies model at about 73% on a construction RFI classification test, highlighting a customer-specific result with a low listed token price.

McCarthy fine-tuned NVIDIA’s Nemotron 3 Nano 4B on more than 100,000 production requests for information, or RFIs. In construction, RFIs are formal questions used to clarify project plans, specifications or site conditions. Nemotron 3 Nano 4B has 3.97 billion parameters.

Palantir’s chart labeled its scoring measure “Balance Accuracy.” The fine-tuned Nemotron scored about 73%, while the untuned version was shown at about 53%. Both entries carried a token price of roughly $0.10 per 1 million tokens. GPT-6 Luna appeared at about 59% and $0.20 per 1 million tokens; Claude Sonnet 5.5 at about 66% and $4; GPT-6 Astra at about 62% and $20 per 1 million tokens.

Those cost figures are approximate readings from the chart, rather than values published in a data table. The chart defines estimated token costs using a 3:1 input-to-output token blend. The update did not itemize hardware, hosting or fine-tuning expenses, so the cost comparison covers token use only.

Palantir described the classifier as a way for project teams to identify emerging risks and assess potential effects on costs and schedules. Rico Bertucci, McCarthy’s vice president of strategic initiatives, said the model was customized to the contractor’s work: “No one else can pick it up, because it is tuned to how we build.”

Palantir did not disclose the test-set size or scoring protocol for the RFI classification exercise. Its chart does not list parameter counts for GPT-6 Luna, Claude Sonnet 5.5 or GPT-6 Astra, so the figures do not establish whether Nemotron is smaller than those systems. The result applies to one classification task and does not demonstrate how the model performs on other construction workflows.

Palantir and NVIDIA announced a sovereign-AI collaboration in September, bringing Nemotron models together with Palantir’s Foundry and Artificial Intelligence Platform. Palantir also gave an annualized estimate of 30,000 hours saved for a separate McCarthy subcontractor-drafting workflow; that estimate is separate from the RFI test.