Skip to Main Content

Burotu

Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

Burotu — Demo video

In one week at the Meta Global AI Developer Hackathon, we turned a large language model into a DNA reader. Meta's open-weight Muse Glimmer 30B had learned from text and images; it had never seen DNA. Burotu teaches it to read DNA methylation from a blood draw: tumours shed DNA into the blood, and that DNA carries methylation marks healthy blood cells don't. Glimmer now says whether there is an unusual signal and which of 23 tissues it looks like. A small adapter turns 3,148 methylation sites into 16 soft tokens in Glimmer's embedding space, the way a vision projector does for images, and Glimmer is fine-tuned with QLoRA following Meta's recipe: 217M trained parameters, 66 minutes on one L40S on Modal. It took six training runs, each dated and documented: v1 (Oct 27) answered the same for everyone; v2 learned to read but memorised donors; v3 and v4 fixed the alarm; v5 beat classic machine learning on the tissue; v6 (Nov 2), trained on 5,728 people from 8 aligned labs, finds 75% of signals at 99.5% specificity on people it never saw, and names the right tissue 58% of the time among 23 at 1% tumour DNA (gradient boosting: 46%). We built the whole data pipeline too: 26,155 public methylation samples refined into one format, 3,832 tumours from 22 public series, 23 tissues. It ships as a product: burotu.com (EN/FR/ES) with a public live demo, where a lab file is read by classic models and Glimmer side by side, with a "why this result" view and a research report for the requesting doctor written by Muse Spark. Burotu is now a reusable DNA-reading engine, online behind one API with open weights on Hugging Face: retrain the adapter on new labels and the same model can follow a treatment, watch for relapse, see which organ is damaged or estimate biological age. It never diagnoses: it is a research prototype trained on simulated plasma and checked on 27 real samples, not a medical device. Everything was built during the hackathon: first commit on Oct 26, 80 commits, about 11,000 lines of Python. Try it live: https://burotu.com/en/demo Weights: https://huggingface.co/zoukmar75/burotu-glimmer

Team