Scale opens Nemotron post-training service to enterprises

Scale AI opened its Post-Training service to enterprises customizing NVIDIA Nemotron open models on October 15 under a collaboration with NVIDIA. The service gives companies a managed way to adapt Nemotron models to specific business tasks.

The offering covers the newly released Nemotron 3.5 Super VL and the wider Nemotron family. NVIDIA describes Super VL as an open, customizable multimodal mixture-of-experts model with 120 billion total parameters and 12 billion active parameters per token. NVIDIA released Nemotron 3.5 Super VL alongside new Bolt models on October 15.

The service is available to any enterprise through Scale’s GenAI Portfolio, according to the company. Scale directs prospective customers to contact its team but does not disclose pricing, standard contract terms or a self-service purchasing option. Financial terms of the NVIDIA collaboration were also not disclosed.

Scale says its workflow starts by evaluating a customer’s current model on the customer’s own tasks. It then builds training and evaluation data, selects a post-training recipe, tunes Nemotron with NVIDIA tools and tests successive versions in the intended agent workflow. Post-training adapts a pretrained model for a narrower task; NVIDIA says its NeMo tools support supervised fine-tuning, low-rank adaptation and reinforcement learning with a customer’s data, tools, policies and workflows.

Scale pointed to work with Skydance, EY and an unnamed global insurer as early enterprise examples. For Skydance’s vendor-credit processing workflow, Scale said tests on real tickets matched the accuracy of the company’s production frontier model at about one-fortieth of the inference cost and with responses five times faster. NVIDIA separately identified the Skydance project as a Nemotron 3.5 Super VL post-training deployment, but the companies did not publish the evaluation data or name the comparison model.

For EY, Scale said a Nemotron model being developed for an M&A due-diligence agent matched the accuracy of EY’s current frontier model across hundreds of questions from held-out data rooms, at about one-eighteenth of the inference cost and more than twice the speed. NVIDIA confirmed the Scale-EY post-training work but did not publish those detailed comparisons.

For the unnamed insurer, Scale said controlled evaluations using synthetic French calls found that a roadside-assistance copilot reached 93.8% fact-extraction accuracy and 95.6% coverage-recommendation accuracy while cutting inference costs by roughly 70% to 75% against the insurer’s production model. Scale said it used supervised fine-tuning and reinforcement learning because the insurer’s call data required legal approval before it could be used for training. The company said approved real-call testing, live-workflow latency measurement and guardrail work remained before production. The opened materials do not include customer-authored validation of the accuracy, speed or cost comparisons.

NVIDIA made Nemotron 3.5 Super VL available through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM on October 15.