Open AI Models Let Companies Build Systems They Actually Own. Here Is What That Looks Like.
NVIDIA's Nemotron family is being shaped by law firms, hospitals and language communities into specialised tools that match frontier accuracy at a fraction of the cost.
Key points
- Harvey, a legal AI firm, post-trained NVIDIA Nemotron 3 Ultra on its own legal benchmark and matched leading closed models at at least 10 times lower cost per run.
- Arcee AI achieved inference costs of roughly 90 cents per million output tokens on Nemotron, approximately 20 times cheaper than comparable closed frontier models.
- H Company's Holotron 3 Nano scored above 76% on OSWorld-Verified, a benchmark of real computer tasks, while matching leading frontier models at a fraction of the cost.
- LangChain tuned its agent framework for Nemotron 3 Ultra with no model retraining and reached top agent accuracy among open models at roughly 10 times lower cost than closed alternatives.
- YTL AI Labs post-trained a Nemotron model for the Malaysian language, making locally customised AI available to Malaysia's developer community.
Most businesses that want AI today face the same quiet problem. The big, powerful models from OpenAI, Google and Anthropic are locked boxes. You can send them questions. You cannot look inside, change what they know, or run them on your own servers without routing your data through someone else's systems.
For a hospital or a law firm, that matters a lot.
NVIDIA has been positioning its Nemotron family of open models, meaning models whose inner workings are fully available for inspection and modification, as the answer to that problem. A blog post published this week by the company lays out what that looks like in practice, with numbers attached.
The legal AI startup Harvey trained Nemotron 3 Ultra on its own internal legal benchmark. The result matched the accuracy of the best closed commercial models on complex legal tasks, at a cost NVIDIA says is at least 10 times lower per run. That is not a rounding error. For a firm running thousands of document reviews a month, the savings compound fast.
Glean built a product called Waldo, a search tool for company data, by pairing Nemotron with a larger closed model. Waldo delivers answers faster and uses fewer tokens, the small units of text that AI models process, which directly cuts computing costs.
Arcee AI pushed further on cost. By running Nemotron on NVIDIA's Blackwell hardware platform, the chip generation that followed the previous Hopper generation, the company brought inference costs down to about 90 cents per million output tokens. Comparable closed frontier models cost roughly 20 times more for the same work.
Not every use case is about law or cost. YTL AI Labs post-trained a Nemotron model specifically for the Malaysian language, creating a customised AI that a global provider would have little commercial reason to build.
Abridge and Heidi Health are both applying customised Nemotron models to clinical documentation, the time-consuming task of turning a doctor-patient conversation into a structured medical note. Getting that wrong carries real consequences, which is exactly why those teams want full visibility into how their model was trained and the ability to fix it when it falls short.
Should businesses ditch closed models entirely?
No. NVIDIA's own framing says open and closed models work best together. A powerful general-purpose model can handle complex planning while a smaller, specialised open model executes specific tasks at lower cost. Think of it as using an expensive specialist for the hard judgment call and a well-trained generalist for the routine work.
The practical argument for open models is not that they beat closed ones on every task today. It is that you own what you build. You can test it against your own data, improve it when it fails, and keep your most sensitive information off third-party servers. For industries where a wrong answer has legal or clinical consequences, that control is not optional.
