Acrylic Surpasses Google, Anthropic, and OpenAI With 31B-Parameter AI Model... Proving 'Sovereign AI' Capabilities

Acrylic, a company specializing in artificial intelligence transformation (AX) infrastructure, has announced that its proprietary AI model specialized for healthcare has outperformed global frontier AI models developed by Google, Anthropic, and OpenAI. By demonstrating the competitiveness of domain-specific AI in the healthcare sector, a field requiring expert knowledge, Acrylic is expected to accelerate its push into the sovereign AI market.

On September 1, Acrylic stated that its healthcare-specialized AI model, ‘AREUM (ALLM).H2 FRANK,’ achieved a perfect score on the physician licensing exam section of ‘KorMedMCQA,’ an open benchmark based on Korea’s national healthcare professional examinations.

In an evaluation targeting a total of 3,009 questions across four professions—physicians, dentists, pharmacists, and nurses—it answered 2,980 questions correctly, achieving an accuracy rate of 99.0%. Based on the total number of correct answers, it outperformed Google’s ‘Gemini 3.1 Pro’ (2,959 questions), Anthropic’s ‘Claude Opus 5’ (2,952 questions), and OpenAI’s ‘GPT-5’ (2,871 questions), respectively.

AREUM.H2 FRANK is a single model with 3.1 billion (31B) parameters. Acrylic conducted comparative evaluations using the same questions, prompts, and scoring methods. Each question was answered five times, with the final answer determined by majority vote, and in the case of the physician test, the model recorded correct responses for all questions in two out of three additional measurement rounds.

Acrylic explained that it validated whether any evaluation questions had been included in the training data, and released the evaluation conditions, code, result ledger, and demo API to ensure the reproducibility of results.

This performance improvement was achieved using its proprietary incremental learning methodology called ‘VICTOR.’ VICTOR selectively adds necessary expert knowledge instead of retraining the entire AI model.

Through this approach, the model’s performance improved by 185 additional questions compared to the base model, and by 174 more questions within three weeks compared to its predecessor ‘ALLM.H,’ which was released in April. Its key feature is the ability to incorporate new expert knowledge or updated regulations and systems without full retraining.

Based on this technology, Acrylic plans to expand the application of its medical AI beyond healthcare professionals to encompass all hospital staff, as well as the bio and pharmaceutical sectors. The strategy is to support not only clinical and medical work, but also research, medical administration, and various other hospital-related work, as well as R&D and clinical work in bio and pharmaceutical companies.

The company is especially focusing on on-premises commercialization. AREUM.H2 FRANK can be directly deployed on GPU servers owned by hospitals or companies, allowing sensitive medical and research data to be processed internally, without being transmitted to external AI services. Acrylic also highlighted the relatively lower inference costs compared to large frontier models as a key advantage.

Acrylic is advancing commercialization through partnerships with major tertiary hospitals in Korea for AI demonstration and deployment, and by utilizing its AI-driven electronic medical record (EMR) platform ‘NADIA-ANE.’ The rollout will begin with hospital work-support services such as cancer screening result summaries, and will gradually be expanded.

Ojjin Park, CEO of Acrylic, said, “The key achievement is that our 31B-parameter model demonstrates expert knowledge comprehension and reasoning abilities that exceed those of global frontier models,” and added, “We will further develop this domain-specific AI so that it can be utilized not only by healthcare professionals but also by various hospital staff, and in the bio and pharmaceutical industries.”

He continued, “Securing data sovereignty in on-premises environments, while continuously adding expert knowledge, is the competitive edge of Acrylic AI. We will strengthen our AX infrastructure competitiveness by expanding the technology proven by AREUM.H2 FRANK into various specialized domains.”

This content was produced with the assistance of AI translation services.