What is MedGemma?
MedGemma is Google's open model family for health AI, released in 4B and 27B sizes and built to understand medical text and images. Ayuvo runs MedGemma 1.5 4B on the device.
On-device AI
Health AI that stays on the phone.
Download Gemma, Qwen or MedGemma once and ask questions with nothing to send. Ayuvo offers MedGemma 1.5 4B, Google's open medical model, as a 2.82 GiB on-device download.
Coming soon to the App StoreComing soon to Google Play
On-device modelsSize · memory
Gemma 4 E2B2.41 GiB · 8 GB RAM
Qwen3 1.7B0.91 GiB · 6 GB RAM
Qwen3 4B2.48 GiB · 8 GB RAM
Qwen3 8B4.55 GiB · 12 GB RAM
Qwen3 14B8.06 GiB · 20 GB RAM
MedGemma 1.5 4B2.82 GiB · 8 GB RAM
Stays on your deviceNo Ayuvo account, no Ayuvo server.
You choose the AIBring your own key, or run a model on the phone.
No ads, no analyticsNothing is tracked, sold or profiled.
Open sourceMIT-licensed. Read every line.
Six models. Zero servers.
Ayuvo can download a model once and run it on the phone itself. Pick one for Coach, Health Records or food analysis, and nothing you ask leaves the phone. A model appears in the pickers only when it is verified and your phone can run it. Smaller phones still see it listed, with the reason it is blocked.
| Model | Reads | Download | Memory | Context | Licence |
|---|---|---|---|---|---|
| Gemma 4 E2B | Text + images | 2.41 GiB | 8 GB | 4,096 | Apache-2.0 |
| Qwen3 1.7B | Text | 0.91 GiB | 6 GB | 4,096 | Apache-2.0 |
| Qwen3 4B | Text | 2.48 GiB | 8 GB | 4,096 | Apache-2.0 |
| Qwen3 8B | Text | 4.55 GiB | 12 GB | 4,096 | Apache-2.0 |
| Qwen3 14B | Text | 8.06 GiB | 20 GB | 4,096 | Apache-2.0 |
| MedGemma 1.5 4B gated | Text + images | 2.82 GiB | 8 GB | 2,048 | Health AI Developer Foundations terms |
- Speech-to-text can run on the phone as well, with the Whisper Base model (about 150 MB) or the native recogniser.
- Apple Intelligence works as an on-device option on supported iPhones. It handles text, not images.
MedGemma
MedGemma 1.5 4B, on your phone.
MedGemma is Google's open model family for medical text and images. It is built to understand medical text, high-dimensional imaging, medical documents and FHIR health records, and Google positions it as a privacy-preserving building block for health AI. Ayuvo runs MedGemma 1.5 4B, a 2.82 GiB LiteRT-LM build, fully on the device, with text and image input.
- Ask Coach about your own diary and records with the conversation processed on the phone
- Needs 8 GB of memory or more, and a 2,048-token context window, shorter than the 4,096 tokens of the other models
- Licensed under Google's Health AI Developer Foundations terms, not Apache-2.0 like Gemma 4 and Qwen3
On-device modelsSize · memory
Gemma 4 E2B2.41 GiB · 8 GB RAM
Qwen3 1.7B0.91 GiB · 6 GB RAM
Qwen3 4B2.48 GiB · 8 GB RAM
Qwen3 8B4.55 GiB · 12 GB RAM
Qwen3 14B8.06 GiB · 20 GB RAM
MedGemma 1.5 4B2.82 GiB · 8 GB RAM
Getting MedGemma
MedGemma is a gated download, so Google asks you to accept its terms first. It takes about five minutes.
- Create a free Hugging Face account.
- Open the MedGemma page and accept Google's Health AI Developer Foundations terms.
- Create an access token with read access.
- Paste the token into Ayuvo's on-device models settings. It is sent only to Hugging Face, and only for gated downloads.
- Tap Download. The file is about 2.82 GiB, checked with a checksum, and can be deleted from the same screen.
What on-device AI sends, and to whom
Running a model is local. Getting one is a download.
Nothing, when it answers
A prompt to an on-device model is processed on the phone. There is no provider, key or server involved.
A download, once
Tapping Download fetches the model file from Hugging Face as a plain request. Hugging Face sees your IP address. For MedGemma your access token goes with it, to Hugging Face only.
Your call, per chat
A conversation pinned to an on-device model never quietly falls back to a cloud model.
Not a medical device. MedGemma is a research model released under Google's Health AI Developer Foundations terms. Its output is preliminary, can be inaccurate and needs independent verification: it is not intended to inform clinical diagnosis, patient management decisions or treatment recommendations. Coach keeps its own rules whichever model answers. It describes, never diagnoses, never suggests starting, stopping or changing a dose, and defers to your prescriber.
MedGemma and Gemma are trademarks of Google. Ayuvo is an independent app and is not affiliated with or endorsed by Google.
Questions about this feature.
Which phones can run it?
MedGemma 1.5 4B needs 8 GB of memory or more. Phones with less still list it, together with the reason it is blocked.
Why does MedGemma need a Hugging Face token?
Google gates the download behind its Health AI Developer Foundations terms. Accept them on Hugging Face, add a free access token in Ayuvo, and the download works. The token is sent only to Hugging Face.
Is MedGemma a medical device?
No. It is a research model whose output is preliminary and needs independent verification. Ayuvo never diagnoses or advises on doses.
Get started
Start with the whole picture.
Free, private, on your device. Bring your own AI key. No account, no subscription, no ads.
Coming soon to the App StoreComing soon to Google Play
Or read the code: Ayuvo is open source