Open AI Models Let Companies Build Systems They Actually Own. Here Is What That Looks Like.

Key points

• Harvey post-trained Nemotron 3 Ultra on its own legal benchmark and matched leading closed models on complex legal tasks at at least 10 times lower cost per run.
• Arcee AI brought inference costs down to roughly 90 cents per million output tokens, around 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, putting locally customised AI in the hands of 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, but you can't look inside, change what they know, or run them on your own servers without routing your data through someone else's infrastructure.

For a hospital or a law firm, that matters enormously.

NVIDIA has been positioning its Nemotron family of open models, meaning models whose weights and workings are fully available for inspection and modification, as the answer. A blog post published this week lays out what that looks like in practice, with numbers.

Harvey, the legal AI startup, post-trained Nemotron 3 Ultra on its own internal legal benchmark and hit frontier-class accuracy on complex legal tasks. Cost came in at least 10 times lower per run than leading closed alternatives. That's not a rounding error: for a firm running thousands of document reviews monthly, 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, cutting computing costs directly.

Arcee AI pushed cost further. Running Nemotron on NVIDIA's Blackwell hardware, the chip generation following Hopper, the company brought inference down to about 90 cents per million output tokens. Closing that gap matters: it's the difference between AI experiments that stay experiments and ones that reach production.

Not every use case is about cost. YTL AI Labs built a Nemotron variant tuned specifically to the Malaysian language, a tool no global provider had much commercial reason to create. Abridge and Heidi Health are both applying customised Nemotron models to clinical documentation, turning doctor-patient conversations into structured medical notes. Getting that wrong carries real consequences, which is exactly why those teams need full visibility into how their model was trained and the ability to fix it when it fails.

We covered the cost pressure pushing enterprises toward this kind of flexibility in our 15 July story on chip demand and enterprise scrutiny.

Should businesses ditch closed models entirely?

No. NVIDIA's own framing says open and closed models work best together. A powerful general-purpose model handles complex planning; 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 routine work.

The practical argument for open models isn't that they beat closed ones on every task today. It's that you own what you build. You can test against your own data, improve it when it fails, keep sensitive information off third-party servers. For industries where a wrong answer has legal or clinical consequences, that control isn't optional.

What I'd watch: whether cost claims hold at genuine enterprise scale, or whether the savings erode once teams add the specialised infrastructure, fine-tuning pipelines and evaluation harnesses that production deployment actually needs.