Scale AI and NVIDIA Team Up to Post-Train Nemotron Models for Enterprises

A new post-training service built around open-weight models

Scale AI says it has had early access to NVIDIA Nemotron 3.5 Super VL and has been post-training it ahead of launch with enterprises including Skydance and EY. Starting with the launch, the service is open to any enterprise through Scale Post-Training, a new offering within the Scale GenAI Portfolio.

The offering is meant to help companies customize leading open-weight models, including Nemotron 3.5 Super VL and the full Nemotron open model family. Scale describes post-training as teaching an existing model a particular job using examples of that job done well, which turns a general-purpose model into a specialist.

For Nemotron, Scale says it combines its evaluation methodology, data and training expertise with NVIDIA post-training tools. The goal is to let enterprises balance cost, performance and control when building agentic systems and workflows that run natively in their own environments.

Why Scale argues for specialized open models

Scale frames the enterprise question as no longer which model is best in general, but which is best for a specific job at a sustainable cost and speed. It notes that inference costs add up quickly at production volume, and that deployments will increasingly mix models of different sizes, modalities, providers, strengths and costs.

Open-weight models are a growing part of that mix, according to Scale, especially when paired with expert post-training for a particular workflow. They can run inside a company's own infrastructure and be tuned to specific tasks, which gives enterprises more control over quality, cost and where their data is processed.

Scale says the process starts with measuring the model a customer runs today on tasks from their own workflow. It then builds training and evaluation data, finds a suitable post-training recipe, post-trains Nemotron with NVIDIA's open tools and tests each version inside the same agent workflow it will use in production. Scale adds that its engineers contribute improvements back to those tools for the open-source community.

Joey Conway, senior director of Nemotron Models and Software at NVIDIA, said: "Open models give enterprises a head start on building with their own data and intelligence." Scale CEO Francis deSouza said the company is showing that with Scale Post-Training, "an open-weight model can match or outperform the models enterprises run today at a fraction of the cost."

Results with Skydance, EY and an insurer

Scale shared three customer examples, all based on its own evaluations.

• Skydance: An agent reviews vendor credit memos, statements and invoices, many of them scanned, and matches them against financial records to find credits the company is owed. On real tickets from the production workflow, Nemotron post-trained by Scale matched the accuracy of the frontier model Skydance runs today, at about 40x lower inference cost and 5x faster response times.
• EY: Scale is building a due diligence agent that answers deal questions by finding and reconciling information across hundreds of files in an M&A virtual data room. Across hundreds of questions from held-out data rooms, the post-trained Nemotron 3.5 Super VL matched the accuracy of EY's current frontier model at about one-eighteenth of the inference cost, with answers returned more than twice as fast.
• Global insurance company: A copilot supports customer service representatives during live French-language roadside assistance calls. It pulls out key case facts, recommends covered benefits and flags when a representative might promise something the policy does not include. In controlled evaluations, Nemotron 3.5 Super VL with Scale Post-Training was up to 31 percentage points more accurate than the company's current production model, at about a quarter of the inference cost.

What it means for enterprises

Scale Post-Training is available now for enterprises that want to customize open-weight models for their own workflows. The company's pitch is that access to model weights alone is not enough: someone has to define what "correct" means for each workflow through a representative evaluation dataset, build the training data and show the system is ready for production.

The results cited are company-reported and tied to specific workflows, so they do not necessarily carry over to other tasks. For enterprises weighing inference costs against control over their data, they point to a route where a smaller, tuned open model replaces a general frontier model for a defined job. Interested companies can contact Scale through its website.