What is Dilr Mira? Clinical extraction, explained

If your team works with clinical documents, you already know the constraint that shapes every tool decision: the records cannot leave your building. Discharge summaries, lab reports, pathology notes and claim forms carry health data, and sending them to a hosted model somewhere else is often the first thing your risk function rules out. So the useful question is not which cloud extraction service is cleverest. It is whether a capable model can run where the data already sits, and return something structured you can actually check.

Dilr Mira, from DILR.AI, is built for exactly that setting. It is a class of small clinical language models that read a document and return structured JSON, on your own hardware, without the document ever going anywhere. This page answers the plain question first, then clears up a confusion the shared name creates: several unrelated things are called Mira, and only one of them is this product. Adoption maturity is still low across the board, with Stanford's 2026 AI Index finding fewer than one in ten organisations have fully scaled AI in any single function, so knowing precisely what a tool is, and what it is not, matters more than the marketing around it.

This is the definitional page. It explains what Dilr Mira is, which Mira you mean, what the current release does today, and where it runs. For the deeper engineering, the full measurement breakdown and the comparison with biomedical models and cloud APIs, it points you to the fuller clinical document extraction guide rather than repeating it here.

This guide is shipped by the team behind Dilr Mira, the private clinical extraction models from DILR.AI. Where a private or schema-matched deployment is the goal, that work runs through DATS, our AI consulting and delivery practice.

What is Dilr Mira?

Dilr Mira is a class of private clinical small language models from DILR.AI, around three billion parameters and built on Qwen2.5-3B-Instruct, that read clinical documents and return source-grounded, schema-valid JSON on the customer's own hardware. It turns a scan, a lab report or a claim form into structured fields a system can use, and it does so where the document already lives, so the data never leaves the building. It is a tool for extraction, not diagnosis.

That last point is the heart of it. Dilr Mira does not decide anything clinical. It reads what a document says and lays it out in a fixed shape, and a person checks the result. The reason it exists is that regulated teams need the structure a language model can produce without accepting the exposure a hosted model would create. Running the model on the customer's own infrastructure removes the transfer entirely, which is the difference between a tool a risk function can approve and one it cannot.

Is Dilr Mira the same as MIRA Pharmaceuticals or the other Mira models?

No. Dilr Mira is unrelated to any of the other things called Mira, and this is worth stating plainly because the name collides in a few places. MIRA Pharmaceuticals, a pharmaceutical company listed on NASDAQ under the ticker MIRA, is a different business entirely. A separate model family also called Mira, published on Hugging Face by a different publisher, has no connection here either. Only one Mira is this product.

If you are looking for this product specifically, the reliable anchor is the publisher and the identifier: it is published by DILR under the model id dilr/Mira-Q2. Because the name is shared, the DILR name and that exact model id are the surest way to reach the right one. Once you know you are in the right place, the rest of this page, and the deeper guides it links to, describe the same product.

What does Dilr Mira do with a clinical document?

Dilr Mira takes an unstructured clinical document and returns structured JSON that follows a fixed schema, with every field grounded in the source text rather than invented. It reads the documents clinical and claims teams handle daily, scans, lab reports, discharge summaries and intake forms, and produces a de-identified record. The patient object holds only age and sex, never names or medical record numbers, keeping the most sensitive identifiers out of any downstream system.

The schema is not fixed to one clinical shape. Mira-Q2 is trained to follow whatever extraction schema you inject in the system prompt, so onboarding a new document type is a matter of a schema file and a few seed examples rather than code changes. The clinical schema it ships with covers ten required fields, among them vitals, labs, medications, diagnoses, procedures and allergies alongside the document type, the encounter and the de-identified patient record. DILR.AI is researching the same schema-agnostic approach for other regulated document types beyond healthcare, though today Mira-Q2 is a clinical extraction model.

What is Mira-Q2, and how does it relate to Dilr Mira?

Dilr Mira is the family, and Mira-Q2 is the current public release within it. When people say Dilr Mira they mean the class of models and the approach; when they need the specific thing to download and run, they mean Mira-Q2, which is the version live today. The two names are not different products, they are the family and its shipped member, in the same way a product line and its current model share a lineage.

Mira-Q2 is open on Hugging Face under the Apache-2.0 licence, and it is deliberately small. It runs on CPU and takes about two gigabytes on disk at four-bit quantisation, with a GPU needed only for speed rather than to run it at all. That size is a design choice, not a limitation: a model that fits comfortably on ordinary hardware is a model a regulated team can actually deploy inside its own perimeter. You can read the model card and download it directly from Hugging Face.

Where does Dilr Mira run, and why on your own hardware?

Dilr Mira runs on the customer's own infrastructure, and that is the point rather than a detail. Because it is small enough to run on CPU, it can sit inside the same environment where the clinical documents already are, so extraction happens without the records being sent to a third party. For any organisation handling health records, that takes the question of a third-party transfer off the table, because there is no transfer to justify.

The legal reason this matters sits in data protection law. UK GDPR treats health data as a special category, which is why casual processing of it is off the table. Article 9(1) puts it directly:

Processing of personal data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs, or trade union membership, and the processing of genetic data, biometric data for the purpose of uniquely identifying a natural person, data concerning health or data concerning a natural person's sex life or sexual orientation shall be prohibited.

That prohibition is not absolute; it is lifted where a condition in Article 9(2) applies, and the duty falls on the organisation processing the data, not on the vendor of an extraction tool. The ICO's guidance on special category data is the reference for how that condition is met. Dilr Mira does not change any of that. What it changes is the exposure: keeping the document on your own hardware means the extraction step adds no new place for the data to go. The fuller treatment of on-premise, CPU-only deployment for regulated data lives in the extraction guide.

How does Dilr Mira handle documents it was not trained on?

Dilr Mira publishes both sides of this openly. On documents that match its training distribution, Mira-Q2 reaches full JSON validity in evaluation; on general real physician prose across many specialties, that figure is about 86 per cent, and the model card states the gap plainly rather than hiding it. The model learned a synthetic generator's patterns very well and is strongest on the lab, discharge and medication document types it saw most.

We think that candour is part of what Dilr Mira is, not a footnote to it. A tool that tells you where it is weaker is a tool you can deploy responsibly, because you know which document types to check hardest and where a human reviewer earns their keep. The full scorecard, with every test set, sample count and confidence interval, is public on the Hugging Face model card and unpacked in the extraction guide; this page stops at the honest headline. The same discipline of publishing what a system can and cannot do runs through how we place AI generally, which is why our DATS methodology and our AI operating model both start from measurement rather than claims. You can browse more of that thinking across our AI strategy writing.

Is Dilr Mira an open model, and can I use it commercially?

Mira-Q2 is genuinely open. It is published on Hugging Face under the Apache-2.0 licence, the same permissive licence as its base model, which allows commercial use. You can download the weights, read the full evaluation files and run the model inside your own environment today, without asking anyone. For a team that wants to test whether small-model extraction is viable on its own documents, that is the fastest honest way to find out.

The download is the easy part; the harder work is the fit: matching the schema to real documents, standing the model up inside a regulated environment, and building the human-review step around it. That work runs through DATS, our consulting and delivery practice, rather than through Mira being sold as a standalone subscription. A useful starting point is a short scoping call, which you can book from our contact page, to see where a model like this actually pays back before anyone commits to a build. If you are weighing the broader consulting question, the enterprise AI consulting guide covers how that engagement works.

What is Mira-3, and is it available yet?

Mira-3 is in development and is not available yet. The DILR.AI models page describes it as the next release, covering multilingual extraction, a zero-shot schema mode and a smaller, faster sibling, but it is explicitly labelled in development, and the current public release remains Mira-Q2. If you are deciding today, Mira-Q2 is the thing you can download and run; Mira-3 is a direction, not a download.

We are deliberate about this distinction because it is easy to blur. Anything you can act on now is Mira-Q2 on Hugging Face; anything described for Mira-3 is a plan we are building towards, and we would rather you knew which was which. When Mira-3 ships, the models page is where it will appear as the latest release. Until then, the honest position is that Mira-Q2 is the product and Mira-3 is the roadmap.

Is Dilr Mira a medical device?

No. Dilr Mira is an extraction tool, not a medical device, and every output it produces is a draft for human review rather than an autonomous clinical decision. It reads a document and structures what is written; it does not interpret, diagnose or advise. A clinician or trained reviewer stays in the loop and remains accountable for how the output is used. Treating it as more than a structured first pass would misread what the model is for.

Who builds Dilr Mira?

Dilr Mira is built by Dilr.ai Ltd, a company registered in Harrow, England, under Company No. 16842656. It sits alongside the wider DILR.AI portfolio, built on one shared engineering approach. The model, its evaluation files and the full scorecard are public on Hugging Face, which is unusual for a commercial clinical model and reflects how the team prefers to work: show the evidence, publish the limits, and let the results carry the argument.

That portfolio includes Cognibl, from DILR.AI, a work-management platform that applies a similar proof-of-done discipline to agent work, and the DATS consulting practice. You can read more about DILR.AI and how it is put together.

Want to go deeper? Read the enterprise AI consulting guide, book an AI placement diagnostic, see how we place AI in regulated systems, or explore the wider AI solutions we build and run.

The same measurement-first logic underpins our AI execution office, where a delivery team runs the operating model day to day rather than leaving it as a slide.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.