Breaking

MedRAX, an AI Reasoning Agent for Chest X-rays, Released With Local Deployment Support

June 30, 2026 at 20:24 EDT

AI Agents
Healthcare
Open Source

A research team from the University of Toronto and affiliated institutes has released MedRAX, an AI agent that interprets chest X-rays (CXR) by combining multiple specialist tools with a multimodal LLM. It handles complex medical queries dynamically without additional training and supports privacy-conscious local deployment.

February 2025 · BoWang Lab · ICML 2025

MedRAX: An Open-Source Reasoning Agent for Chest X-rays

A single agent orchestrates specialized tools — visual QA, segmentation, grounding, report generation, and disease classification — for multi-step clinical reasoning. It runs locally with quantized LLMs, and tops a new benchmark of 2,500 expert-verified queries.

63.1%

ChestAgentBench accuracy — leads general-purpose & specialized models

90.35%

SLAKE visual QA accuracy

79.1%

mF1-14 for MIMIC-CXR report generation

ChestAgentBench accuracy

2,500 complex queries from 675 expert-verified clinical cases

63.1%

MedRAX

57.9%

Llama-3.2-90B

56.4%

GPT-4o

39.5%

CheXagent

28.7%

LLaVA-Med

One agent, many specialized tools

Visual QA
CheXagent · LLaVA-Med

→

Segmentation + Grounding
MedSAM · Maira-2

→

Classification + Reports
DenseNet-121 · SwinV2

→

LangGraph resolves
conflicting outputs

Supports GPT-4o plus local LLMs via Ollama / LM Studio with 8-bit / 4-bit quantization — enabling offline, privacy-preserving deployment. No additional training required.

Strengths

Robustness from integrating many tools
Practical local / offline deployment
Transparent multi-step reasoning

Limitations

Higher resource use than a single model
Manual setup for some tools; benchmark-heavy evaluation
Physician judgment still essential — no self-diagnosis. Grok-4 reportedly hit 72.8%; MedRAX2 in development.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.