Aigen Sciences and Korea University introduced Meerkat2, a medical AI language model for Korean clinical settings, at EMNLP 2026 in Budapest. The model is available in 14 billion, 32 billion, and 30 billion parameter versions. It cuts answer length by about 63% while improving accuracy on Korean medical licensing questions.
Training on Korean law
Researchers collected 12 Korean medical laws and built 18,700 question-answer pairs from the provisions. Meerkat2 scored 98.7 points on legal accuracy, above the 79.0 points from the previous method. It answered 84.6% of medical law questions on the Korean licensing exam, compared with 76.9% for GPT-5.1.
Company background
Aigen Sciences was founded in 2021 by Kang Jae-woo, a Korea University computer science professor. Its 2024 Meerkat-7B model was the first sub-7 billion parameter model to pass the USMLE. The company also supplies Meerkat models to Kakao Healthcare's Dr.LIKE service.
How the model works
The model uses a training method called Verbosity-Decoupled Distillation. It learns detailed reasoning from larger models without copying their long answer style. A 30 billion parameter mixture-of-experts version activates about 3 billion parameters per query and matches a 235 billion parameter model on English licensing questions.
Meerkat2 is available for free on Hugging Face. Aigen Sciences also offers an enterprise version with higher accuracy for clinical institutions. The company is applying the same language model technology to its drug discovery pipeline.