Does medical model architecture or training data drive performance gains?

The MedGemma technical report argues that its models' medical competence comes largely from domain data in pretraining and post-training, not from a new architecture. MedGemma 4B and 27B are built on Gemma 3, and the report's own evaluations show improvements over the base models of 2.6-10% on medical multimodal question answering, 15.5-18.1% on chest X-ray finding classification and 10.8% on agentic evaluations, all on out-of-distribution tasks. Fine-tuning cuts errors in electronic health record information retrieval by 50%, and for pneumothorax and histopathology patch typing the report says it reaches "comparable performance to existing specialized state-of-the-art methods." The authors credit "optimized incorporation of domain specific data for both pre-training and post-training," and say the gains hold "across all benchmarks evaluated."

The mechanism is data selection and mixing, with the architecture held at Gemma 3. The text models learn from responses and logits sampled from a large instruction-tuned teacher over medical QA sets such as MedQA and PubMedQA, plus about 200,000 synthetic questions. The image data is curated: PathVQA and MedVQA are dropped after the authors identified "potential data quality issues," and internal collections add 184,852 retinal fundus images, 51,049 dermatology images and about 32.5 million histopathology patch-text pairs. The vision encoder, SigLIP-400M, is fine-tuned on over 33M medical image-text pairs, with its original WebLI data kept and medical data mixed in at 2% weight so that its general performance survives.

Against the nearest notes, the report fits a domain-investment reading of medicine but cannot test which capability moves. Its investment is in medical knowledge data, which is consistent with Does medical AI need knowledge or reasoning more?, though the report never splits its scores into knowledge and reasoning. Its retention lever differs from Can we solve modality competition through architectural design?: MedGemma stays dense and protects general vision performance with a data mixture, where that note locates the fix in architectural capacity. Training medical knowledge into the weights is one point in the flexibility-cost trade of How do knowledge injection methods trade off flexibility and cost?, and the report's reasons for choosing it (a frozen model, cost sensitivity, local operation) come without cost figures. Benchmark scores also do not test the barriers in Can clinical experts teach LLMs to annotate complex medical concepts?, which were serious enough there to require changes to the extraction workflow.

What the excerpt does not establish matters most for the title. Every figure is the report's own benchmark result, not an independent replication, and the report itself says that "automated benchmarks represent only the first step towards validating real-world utility" and that some benchmarks are near saturation. The excerpt has no clinician study or deployment data, does not name the evaluation sets behind most of the ranges, and does not reproduce the comparison with "similar-sized generative models." It also has no ablation separating pretraining data, post-training and vision-encoder fine-tuning: the 4B model used all the listed steps, while the text-only 27B used post-training alone. The base comparison holds the architecture fixed, which is why the report can credit training, but "largely" is the report's word and the excerpt cannot test it. The defensible reading is narrower than the report's framing: domain data moved a Gemma 3 base on these benchmarks, which is evidence about benchmarks rather than about clinical usefulness.