hjLabs AI Playground
Starting the on-device runtime…
MedGemma is Google’s open medical model family — 4B multimodal, 27B text, and the newer 1.5-4B that adds 3D CT and MRI volumes, whole-slide pathology and EHR understanding. It is the most downloaded medical model on Hugging Face, and it is the one thing this playground gets asked for most. It cannot run here, for three reasons that are each independently fatal: every Google MedGemma repo is `gated: auto`, so an unauthenticated browser fetch is refused with a 401 before a single weight byte; no export in the transformers.js layout exists anywhere on the Hub; and even the small 4B variant is about 2.5 GB at 4-bit, past what a tab should download. This page shows the checks, quotes the licence it ships under, and points at the models that do run in this browser and at the one command that runs MedGemma on your own machine.
medgemma browserrun medgemma locallymedgemma onnxmedgemma webgpumedgemma transformers.jsmedgemma ggufmedgemma 4b requirementsmedgemma llama.cppis medgemma freemedgemma license
No. Every google/medgemma-* repository is gated (`gated: auto` in the Hub API), so a browser fetch with no Hugging Face token gets HTTP 401 on config.json — the request fails before any weights are reached. Even if it were ungated, transformers.js needs an ONNX export laid out as onnx/model.onnx plus a matching config, and no such export exists for any MedGemma variant. Google publishes none; the third-party conversions are onnxruntime-genai format, 5–57 GB, and the only 4B one inherits the gate.
Four, plus an image encoder. medgemma-4b-it is 4.3B parameters and multimodal (images and text). medgemma-1.5-4b-it is the same size and adds 3D CT/MRI volumes, whole-slide histopathology, longitudinal chest X-ray comparison and EHR understanding. medgemma-27b-it is 28.8B and multimodal; medgemma-27b-text-it is 27.0B and text-only. MedSigLIP-448 is the 878M image-text encoder the multimodal ones use.
No. It is Health AI Developer Foundations (HAI-DEF), which Google tags as license: other. HAI-DEF does permit hosting it as a service, contrary to the common assumption — but it requires you to pass the use restrictions on as an enforceable term to every user, to seek health regulatory authorisation where applicable, and it makes you indemnify Google, uncapped, for injury arising from your use, under California law in Santa Clara County. Read it before you build on it.
llama.cpp with a GGUF build is the shortest path. unsloth/medgemma-4b-it-GGUF is not gated, so llama-server -hf unsloth/medgemma-4b-it-GGUF:Q4_K_M pulls a 2.49 GB file and serves an OpenAI-compatible endpoint on localhost:8080. Add the 851 MB mmproj file if you want it to accept images. Roughly 4 GB of VRAM or unified memory for Q4_K_M text-only, about 5 GB with vision.
It is not a medical device and its output is not a diagnosis. Google’s own model card is explicit: "The outputs generated by MedGemma are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications." Nothing on this site is medical advice, and no model here — MedGemma included — can tell you whether something is wrong with you. A clinician can.
Text models, and they work well: biomedical entity extraction across 84 clinical types at 67 MB, clinical PHI de-identification, drug and disease taggers, and UMLS-aligned embedders that treat "MI" and "myocardial infarction" as one concept. There is also a DICOM viewer that opens a .dcm file, shows the patient identifiers hiding in its header and strips them, with no model at all. There is no medical image classifier here, because none exists that this runtime can load and that is licensed for a commercial site.
No. The model weights are downloaded from Hugging Face to your browser once, and every inference after that runs on your own GPU through WebGPU. Your text, images and audio are never sent anywhere — hjLabs.in has no inference server and no way to see your input.
That depends on the individual model, not on this site. Each model page links to its Hugging Face repository, where the licence is stated — most of the catalog is Apache-2.0 or MIT, a few are research-only. We add nothing on top: hjLabs.in claims no rights over anything you produce here.