Lunit open-sources LearningUnit 2.0, a 30-billion-parameter medical model built for hospital servers
The Korean model has 29.56 billion total and 5.34 billion active parameters under Apache 2.0 on Hugging Face; Lunit's performance claims, including parity with GPT-5.2 on its benchmark set, are its own and were reported by Korean press.
Why it matters
Lunit built the model as the lead of a roughly ten-month, 23-organisation Korean government project, and says it was tested at nine medical institutions. It targets hospital workflow rather than chat: clinical decision support, emergency-room triage, multidisciplinary care, adverse-drug-reaction causality and insurance review. A model small enough to run inside a hospital addresses the rule that patient data should not leave the institution. The company says it beat the 120-billion-parameter GPT-OSS on its medical benchmarks. Watch for independent evaluations and real clinical deployments.
Key facts
- Model L1-30B-A5B: 29.56B total and 5.34B active parameters, sparse mixture of experts
- 131,072-token context; Apache 2.0 licence on Hugging Face
- Built in a Ministry of Science and ICT project with 23 organisations over about 10 months
- Piloted at nine medical institutions (Lunit, via News1)
- Lunit says it beats the 120B GPT-OSS and is comparable to GPT-5.2 on 17 benchmark items