NVIDIA and Palantir fine-tune a 30B Nemotron model that beats an 18x-larger model on NVIDIA's own supply chain

NVIDIA and Palantir have applied a fine-tuned 30-billion-parameter Nemotron model to NVIDIA's own supply chain allocation decisions, publishing results showing the smaller specialized model far outperforming NVIDIA's much larger general-purpose Nemotron model on the same task.

What's new

In a technical blog post dated Sep 10, 2026, NVIDIA describes how "the NVIDIA supply chain operations team worked with Palantir to create a unified view of every input to a material allocation decision." The team then post-trained Nemotron 3.5 Lightning, a 30B model, on captured allocation decisions, planner rationales, and outcomes, and compared it against Nemotron 3 Ultra, a much larger general-purpose model.

The results are stark: "the post-trained Lightning model reached 86.7% allocation-decision accuracy. Ultra reached 55.5% and base Lightning 17.5%. That puts the post-trained model 31.2 percentage points ahead of Ultra and 69.2 ahead of its own base model." NVIDIA frames this as the 30B specialist beating "a general-purpose model that is more than an order of magnitude larger."

The fine-tuning itself was fast and cheap by design. NVIDIA's NeMo AutoModel "trains a small set of LoRA adapter parameters while the base weights stay frozen, which cuts training time, memory requirements, and checkpoint size," and the actual training run completed using "2x NVIDIA B200 GPUs in minutes." Two other NeMo tools did the data prep: NeMo Anonymizer stripped sensitive information from the training data, and NeMo Data Designer expanded the available training examples. On the Palantir side, "Palantir Autopilot manages the lifecycle end to end, launching each job from Ontology data" — meaning the entire pipeline, from raw operational data to a deployed fine-tuned model, runs inside Palantir's governed data platform rather than a separate ad hoc training environment.

Context

This is a real-world deployment on NVIDIA's own multi-billion-dollar chip supply chain, not a benchmark exercise on public data. Supply allocation is a problem NVIDIA has particular reason to care about given the scale of demand for its GPUs, and the post makes clear the underlying quantitative optimization (a mixed-integer linear program) is solved separately by NVIDIA's cuOpt — the Nemotron model's job is to capture qualitative planner judgment (things like weather, geopolitical risk, or supplier emails) that a pure solver misses, not to replace the optimizer itself.

Why it matters

The headline number — a 30B model beating a model over 18 times its size by more than 31 points of accuracy — is a concrete data point in the broader argument that domain-specific fine-tuning on an organization's own operational history can beat throwing a bigger general-purpose model at the same problem, at a fraction of the compute cost (minutes on two GPUs, versus running a much larger model at inference time indefinitely). For NVIDIA specifically, showcasing this on its own supply chain doubles as a reference deployment it can point enterprise customers to when selling the same NeMo/Palantir Foundry combination for their own high-stakes allocation, planning, or operations decisions.