Final chapter of the Sprint DGX series. We close out the grant with a dental vision model that’s trained, validated, and ready to live inside the practice, with our sights already set on the next clinic.
We started these 60 days with one concrete idea and the chance to test it at scale: that the vision brain of a real dental product can be trained at data-center scale and still live entirely inside the clinic. Today, as we close out the NVIDIA Innovation Lab grant, that idea is no longer a hypothesis. It’s a model that’s trained, measured, and ready to plug into a system that sees patients every day. This is the closing chapter: what’s left standing when the GPUs power down.
What we delivered
A vision-language model specialized in dentistry, starting from Gemma 4 31B, one of the most capable open models in its size class, and fine-tuned on our own clinical data on a node of eight NVIDIA H100 GPUs.
It’s not a chatbot you slap a photo onto. It’s a model that looks at the radiograph and reasons about it in the same step: panoramics, periapicals, intraoral photos. It interprets the image directly, not a description of the image produced by some other system. And it does so equally well in Spanish and English, verified with our own purpose-built tests, not as a second-rate “translation mode” but as genuine parity. For a Spanish-speaking clinical market, that’s no small thing.
On a public dental-knowledge exam (MMOral-Bench), the specialized model gains +6.72 percentage points in English and +7.13 in Spanish over the base model, a highly competitive result for its size class, ahead of generalist models that are far larger and more expensive. And it responds fast enough to assist the professional during the consultation, not to keep them waiting.
Why we decided it should live in the clinic
This, for us, is the part that truly sets the product apart. The model is designed and validated to run on-premise, inside the clinic, on an NVIDIA DGX Spark, NVIDIA’s desktop AI computer, with 128 GB of unified memory, as the minimum configuration, and up from there. There’s plenty of memory to load the full model comfortably, and deployment on Spark has already been validated.
What does that mean in practice? That not a single patient image leaves the building. There’s no health data traveling to a third-party server, no dependence on a connection, no cloud provider sitting in the middle of the conversation between the dentist and their information. In a sector where the data is especially sensitive, and the regulation especially demanding, “it stays home” isn’t a technical concession: it’s an advantage from the outset.
This is not a demo
The product this layer will plug into is already in production at a real clinic, serving several thousand patients and managing hundreds of thousands of images reliably. The sprint didn’t build a prototype to show off in a video: it trained the hardest layer, the intelligence layer, of a system that already works, so that what the professional already uses every day is, simply, better.
That’s why what remains at the close of day 60 isn’t an experiment to file away. It’s a foundation to keep building on: a deployable dental checkpoint, multimodal, bilingual, ready to measurably improve the quality of a real product.
Where it’s headed
The goal hasn’t changed since day one: to put in the professional’s hands an aid that sees what matters, sooner and better, without asking them to give up control of their data or their practice.
These 60 days were the most demanding part, training the model at scale and proving it fits entirely inside the clinic. With that groundwork solved, the next step is bringing it to more clinics: integrating it into every practice, adapting it to more equipment, and continuing to fine-tune it with real-world use. The grant closes one stage and, above all, enables the next.
Because underneath all this there’s a very concrete purpose. The mouth is one of the best places in the body to catch things that matter in time: a lesion that’s changing, a pattern that doesn’t fit, a detail worth a second look. What we built is that extra pair of eyes, tireless and consistent, that helps the professional miss nothing.
Closing the series
This series has documented one specific, well-defined sprint: 60 days, a node of eight H100 GPUs, a small team, and the ambition to train a dental AI that lives where it has to live. What remains at the close is what counts, a model that sees, reasons in two languages, and is designed to run on-premise, ready to be integrated into a product that already cares for real patients.
Thank you for following us chapter by chapter. This isn’t an ending, but a starting point with plenty of ground ahead.
If you work in a dental group, invest in digital health, or are simply curious about how to build clinical AI that genuinely respects privacy, we’d love to show you. Reach out.
This closes the Sprint DGX series. Thanks for following it this far, anyone who wants to dig into a specific chapter, or just talk about what got built, is welcome to reach out.
FAQ
What hardware will it run on? The deployment target is an NVIDIA DGX Spark, 128 GB of unified memory, as a minimum, and any more powerful machine. Plenty of memory to load the full model, and deployment on Spark has already been validated: data-center-class power in a box that fits in the practice, not in a remote data center.
Does patient data really never leave the clinic? That’s the whole idea behind the design. The model is meant to run inside the center itself, so that images are analyzed locally and never travel to any external server, with no dependence on the cloud or a permanent connection.
Is it available for my clinic? The base system is in production and the integration of this new vision layer is rolling out in phases. We’re open to conversations with interested clinics and partners, reach out and we’ll show you.
