Ai Engineering 2 min read

Alibaba Open-Sources RADAR, a Medical AI That Beat Most Radiologists on Abdominal CT Scans

Alibaba's Damo Academy open-sourced RADAR, a medical AI trained on 420,000+ CT exams that detects nearly 150 abdominal conditions with an AUC of 0.913, outperforming 23 of 26 radiologists in comparative testing.

Alibaba’s Damo Academy open-sourced Damo Radar on September 20, a vision-language model that analyzes contrast-enhanced abdominal CT scans and detects nearly 150 abdominal conditions including cancers, per South China Morning Post. The model was trained on over 420,000 contrast-enhanced abdominal CT examinations covering 18 abdominal organs with roughly 15 million anatomy-aware annotations, and achieved an average AUC of 0.913 across 146 clinical findings, tested on nearly 40,000 real-world examinations. In comparative testing, it outperformed 23 of 26 radiologists. The code and model are on GitHub, and the underlying research was published in Science.

Why the Training Scale Matters More Than the Benchmark

Medical AI models routinely post strong benchmarks on curated datasets and fail on real-world distributions. RADAR’s 420,000-examination training corpus with 15 million anatomy-aware annotations is notable because it was built from real clinical imaging across 18 organs, not a curated challenge set. The 40,000-examination test set is also drawn from real-world data rather than a held-out benchmark split, which means the 0.913 AUC reflects the distribution shift every deployed medical model faces. That distinction is the difference between a paper result and a deployable capability.

The Open-Source Decision Is the Strategic One

Alibaba open-sourcing RADAR follows its open-weight strategy for Qwen and positions DAMO Academy as the leading publisher of open medical AI models. The licensing terms matter enormously: open-source medical AI that hospitals can run on-premise serves a fundamentally different market than cloud-only models, and in markets where patient data cannot leave the hospital network (which is most of them), on-premise is the only deployment option. Alibaba also has COCA, its colorectal cancer detection model, in its medical AI portfolio, suggesting a deliberate strategy of open-sourcing horizontal medical AI capabilities.

What to Watch

Three things determine whether RADAR reaches patients. First, regulatory clearance: open-sourcing the weights is not FDA or NMPA approval, and medical imaging AI requires clinical validation per jurisdiction. Second, the radiologist community’s response: the model beat 23 of 26 radiologists on benchmarks, but real clinical adoption depends on integration into existing PACS workflows and radiologist trust. Third, whether other Chinese tech companies follow Alibaba’s open-source medical AI strategy, which would create a parallel ecosystem to the US cloud-dominated medical AI market. The medical AI adoption concerns the FT reported this week, citing limited clinical data beyond diagnostics, are the counterweight to the benchmark enthusiasm.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

Build a Fast Multilingual OCR with Nemotron-OCR-v2

Learn how to deploy NVIDIA Nemotron-OCR-v2 for high-speed document extraction across six languages using synthetic data and GPU acceleration.

Ai Engineering

Qwen4-Max: Alibaba's 2.4T closed model cuts prices and claims a 62% cheaper task, but trails US rivals on independent tests

Alibaba released Qwen4-Max on November 16 with a new Gated DeltaNet plus sparse-attention architecture, 71B active parameters, 1M context and $1.50/$4.50 pricing. Its own Work360 benchmark shows a 62% lower cost per task; Artificial Analysis puts it at 43, behind Kimi K3.1 Pro.

Ai Engineering

llama.cpp 1.0.0: semantic versioning for libllama and ggml, FP8 KV caches and a safer llama-server

llama.cpp and ggml shipped version 1.0.0 on November 13, three and a half years after the first commit. The release commits both libraries to major-version bumps for breaking C API and ABI changes, removes deprecated functions, and tightens llama-server's network defaults. Here is what changes for anyone who builds on it.

Ai Engineering

A Dermatologist's 10-Year-Old CNN Beat GPT-6 Astra and Fable 5.1 at Skin Cancer

Dr Magnus Lynch's October benchmark study fine-tuned a 2015 Inception v3 on 68,011 dermoscopic images on a single H100 and beat GPT-6 Astra and Claude Fable 5.1 at diagnosing 200 biopsy-proven skin lesions: 62% versus 42% first-choice accuracy, and 13 of 30 melanomas versus 8 and 11.

Ai Engineering

DeepSeek Open-Sourced Its Production Inference Engine, and Huawei Gets Equal Billing

DeepSeek released DeepInfer on October 20 under Apache 2.0: the inference framework behind its production services, with NVIDIA SM100 and Huawei Ascend 950 support given equal footing, optimized kernels, and speculative decoding built around confidence scheduling.