How Leaders in Supply Chain Used a Specialised 30B Model to Outperform a Larger One on Allocation Decisions
On a bounded allocation task, NVIDIA’s specialized 30B model can outperform a general-purpose model that is more than an order of magnitude larger.
NVIDIA’s post-trained Nemotron 3.5 Lightning, a 30-billion-parameter model with roughly 3 billion active per forward pass, reached 86.7% allocation-decision accuracy on the development benchmark. Nemotron 3 Ultra reached 55.5% and base Lightning 17.5%, a lead of 31.2 and 69.2 percentage points. The LoRA run finished on 2x NVIDIA B200 GPUs in minutes.
- 86.7%
- allocation-decision accuracy on the development benchmark
- +31.2 pts
- over Nemotron 3 Ultra (55.5%)
- +69.2 pts
- over base Nemotron 3.5 Lightning (17.5%)
- 58.6% vs 42.0%
- balanced accuracy, against Nemotron 3 Ultra
The problem
NVIDIA must decide what and how much material to allocate to each manufacturing site, a problem it reworks manually every week. Back-testing historical allocation decisions revealed a human factor that cuOpt wasn’t capturing: planners were working from information the solver could not see, such as partner emails, weather forecasts, geopolitical events and supplier debrief transcripts.
What they did
NVIDIA and Palantir built a workflow that captures the allocation decision, the rationale behind it, the expected result and the actual outcome. Nemotron 3.5 Lightning has 30 billion parameters and roughly 3 billion active per forward pass.
The result
Macro-F1 was 57.5% against Ultra’s 39.5%. The LoRA run finished on 2x NVIDIA B200 GPUs in minutes.
Worth noting
NVIDIA writes: “This doesn’t mean the smaller model is more capable overall. Its gains are concentrated in the domain it was post-trained on. Future production risk forecasting remained difficult despite fine tuning.”
Figures come from the source below. Anything from another place is attributed in the text.
Source: NVIDIA Technical Blog: from wafer-out to first token ↗