September 19, 2026
Medical AI is faster to launch on LLMs: setup took 8 person-hours
A Webiomed study dated September 15, 2026 found that LLM tuning took 8 person-hours versus 340 for the baseline specialized model. Across 171 clinical cases, YandexGPT-8B and Webiomed SC2.0 achieved 67% Top-3 accuracy, while Webiomed SC2.1 reached 78% on 18 million records.
An LLM gets a medical task to its first result faster. Running continuously under high load has different economics: YandexGPT-8B is estimated at 61.4 thousand rubles per month, while Webiomed SC costs 3.8 thousand rubles. The authors recommend LLMs for prototyping and specialized models for regular workloads.
Infrastructure is growing. On September 16, Moscow Oblast signed two contracts for 12 AI platforms for digital projects worth 704.2 million rubles. Hightech-Integration will supply 10 systems for 577.5 million rubles, while Russian Digital Company will supply another 2 systems for 126.7 million. The platforms will receive at least four NVIDIA H200 accelerators or equivalents, and two systems will be equipped with eight NVIDIA H100 chips or equivalents.
Clinical tasks. SamSMU is developing a system that assesses disease probabilities from heart and lung auscultation recordings. The project was commissioned by FSAI Digital Industrial Technologies for 48 million rubles. Through December, the developers are refining classification, detection of extraneous noise, merging recordings from different points, MIS integration, and the dataset.
The market is expected to grow. Wissen Research estimates the global market for virtual medical AI assistants at $2.6 billion in 2026 and projects $12.3 billion by 2031, with compound annual growth of 36.5%.
Platform deliveries must be completed by the end of 2026, while the SamSMU system is planned for completion in December.
