Google announces new medical AI model that outperforms GPT-4

Last month, Google announced a new family of artificial intelligence (AI) models aimed at the medical field. These AI models, known as Med-Gemini, are not yet available to the public, but the tech giant has published a preprint version of a research paper outlining their capabilities and techniques.

The company claims that its AI model outperforms the GPT-4 model in benchmark tests. One of the distinguishing characteristics of this AI model is its long-context capabilities, which enable it to process and interpret health data and research publications. The research results are currently in pre-print stage and published on arXiv, his open access internet repository for academic papers.

“I'm very excited about the potential of these models to not only help clinicians provide better care, but also help patients gain a deeper understanding of their medical conditions. In my opinion, health “AI for care will be one of the most impactful application domains for AI,” Jeff Dean, principal scientist at Google DeepMind and Google Research, said of X.

Gemini 1.0 and 1.5 LLM serve as the foundation for the Med-Gemini AI model. There are four versions in total: Med-Gemini-L 1.0, Med-Gemini-M 1.0, Med-Gemini-M 1.5, and Med-Gemini-S 1.0. Each multimodal model can produce text, images, and video output.

The model is powered by online search, powered by self-training to make the model “more factual, reliable, and nuanced” when presenting results for complex clinical reasoning tasks. is combined with

The company further states that its AI models are optimized for increased speed when processing long contexts. High-quality long-context processing allows chatbots to provide more accurate and pinpoint answers even when questions are not fully queried or when large volumes of medical records need to be processed.

According to Google statistics, the Med-Gemini AI model outperformed OpenAI's GPT-4 model in the GeneTuring dataset on text-based inference tasks. Med-Gemini-L 1.0 also achieved an accuracy of 91.1 percent in MedQA (USMLE), outperforming the previous generation Med-PaLM 2 by 4.5 percent. In particular, AI models are not available for public release or beta testing. The company may make further improvements to the model before releasing it to the public.