Med-Gemini

// Overview

Med-Gemini is a family of medical research models from Google, built on the multimodal Gemini models and specialized for medicine. Across 14 medical benchmarks, the paper reports new state-of-the-art results on 10 of them, and surpassing the GPT-4 model family on every benchmark where a direct comparison was possible.

On the MedQA (USMLE) benchmark, the best Med-Gemini model reached 91.1% accuracy using an uncertainty-guided search strategy — 4.6 points above the prior best, Med-PaLM 2. Med-Gemini is a research effort described in a 2024 paper rather than a shipping product, and its multimodal reach spans text, medical images, video, and electronic health records.

// Benchmarks

1. MedQA (USMLE)91.1%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• State-of-the-art accuracy on many medical question-answering benchmarks
• Multimodal — handles medical text, images, and video
• Uncertainty-guided search that can defer to web search when confidence is low

// Best for

• Research into multimodal clinical reasoning and diagnosis support
• Benchmarking medical question answering against expert-level baselines
• Exploring long-context reasoning over electronic health records

// Med-Gemini / Med-PaLM (medical) — every version

The full lineage of the Med-Gemini / Med-PaLM (medical) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.