MedMLX, a collection of 21 pretrained medical imaging models rewritten to run natively in MLX on Apple Silicon, reproduced the outputs of their original implementations in 211 of 212 comparisons against an Intel CPU and an NVIDIA RTX 4090, and with a memory-saving setting 19 of the 21 models ran in under 16 GB of memory on a 64 GB Apple M1 Max. The project comes from the AI-IR lab at the University of Texas Medical Branch, whose Division of Vascular and Interventional Radiology rebuilt nnU-Net, TotalSegmentator, MedSAM2 and SAM 3 among 16 architecture families, and published the benchmark as a medRxiv preprint on 29 October 2026.
"MedMLX enables on-device AI with established medical imaging models on Apple Silicon computers rather than specialized NVIDIA hardware," the authors Joseph Sandoval, Alfredo Sandoval Jr. and Arsalan Saleem conclude in the preprint. They add that the models agree closely with the original software and that most need less than 16 GB of memory.
Medical imaging frameworks such as nnU-Net and MONAI are developed primarily for NVIDIA GPUs and CUDA, and a single study may run several models over the same volumetric images, making memory capacity a practical constraint. The researchers point out that macOS runs on about one in six desktop computers worldwide and that a third of software developers responding to a 2025 survey used it for their work, so on-device inference could broaden access to imaging AI research for clinical researchers without specialized hardware. Running on device also means imaging data need not be transferred to shared or cloud computing services.
Before MedMLX, a native SAM 2.1 implementation and medical tools running on PyTorch's Metal backend supported local analysis, but extending that support across tasks required preserving the behavior of each released model.
MedMLX reimplements published models in MLX, Apple's machine learning framework, and the inference path uses no PyTorch or CUDA. Each package combines a pretrained network with the preprocessing and postprocessing of its original implementation, with segmentation masks mapped back to the input image grid, and model-specific operators reproducing the original rounding and accumulation behavior where that affects output agreement. A shared Python API, command-line interface and local Model Context Protocol server expose the models for listing, describing arguments, checking inputs and weights, and running inference, so compatible clients including AI coding assistants can drive the same model operations. Disabling the MLX cache is the memory-saving setting, and optional higher-precision arithmetic profiles for renal SegResNet and SynthSeg 2.0 let those models produce identical masks across all three devices. MedGemma runs through a provisional bridge to a separate MedVLM environment, with checkpoint inference and numerical accuracy left unqualified.
Across public cases from 11 imaging datasets, MLX outputs were identical to or within tolerance of the original outputs in 211 of 212 comparisons, covering 20 of the 21 models plus the FleXray ensemble and conditioned LesionDiffusion. Agreement required a Dice score of at least 0.99 for every mask label and cosine similarity of at least 0.9999 for images and features. The single comparison outside tolerance was SynthSeg 2.0 on BRATS_004, where the original Intel CPU and RTX 4090 outputs also differed beyond tolerance from each other. BioMedParse v2 on a 63-slice AMOS22 CT produced identical spleen and right-kidney labels between FP32 MLX on the M1 Max GPU and the original PyTorch implementation on the same computer's CPU. With matched random noise, LesionDiffusion produced identical lesion masks, with mean absolute differences below 1 HU inside and outside the lesion masks, although local differences outside the masks reached 428 HU. The memory-saving setting reduced memory use by a median of 49% and increased processing time by a median of 13%.
The authors present MedMLX as research software that is not a medical device and not for clinical use, with outputs requiring human review, and the AI-IR lab lists it under accessible AI research aimed at lower-cost imaging tools. The release also includes Reason-CT with nodule tools, combining model-selected detector and MedSAM2 calls, spatially grounded masks and a local CT review page, and the preprint was funded by the John Sealy Memorial Endowment.