| --- |
| license: other |
| license_name: medpodgp-eula |
| license_link: LICENSE.md |
| language: |
| - en |
| tags: |
| - medical |
| - clinical |
| - healthcare |
| - on-device |
| - local-ai |
| - speech-to-text |
| - medical-imaging |
| - desktop-app |
| pretty_name: MedPod GP — On-Device Medical AI Assistant |
| --- |
| |
| # MedPod GP — On-Device Medical AI Assistant |
|
|
| **MedPod GP** is a **desktop clinical AI assistant** for general practitioners that |
| runs **entirely on your own computer**. Record and transcribe consultations, ask |
| clinical questions, read medical images, search clinical guidelines, and draft |
| structured notes for the record — with **no cloud account, no telemetry, and no |
| network calls for inference**. Patient information never leaves the computer. |
|
|
| A product of **CloudKites AI Lab (CloudKites Pty Ltd)**, Sydney, Australia. |
|
|
| > 💚 **Free to download, use, and share.** MedPod GP is distributed **free of |
| > charge**. You may use it and **redistribute the complete, unmodified official |
| > package** for free (no selling, no paid bundling). See **License** below. |
|
|
| > ⚕️ **Clinical documentation and decision-support aid — not a medical device.** |
| > MedPod GP does not diagnose, treat, or replace professional judgement. All AI |
| > output must be reviewed and verified by a qualified clinician. See **Medical |
| > disclaimer & intended use** below. |
|
|
| > 🎉 **Latest release: MedPod GP 3.2.1** — native on **macOS, Windows and Linux**. New in 3.2.1: |
| > **speech recognition entirely on the NVIDIA card** (75 s of speech ready about 2 s after |
| > stop on an RTX laptop) and **far less memory on Windows** (a recording 3.5 GB instead of 9.6, |
| > an X-ray 4.9 GB instead of 7.5). New in 3.2: |
| > **Windows uses the NVIDIA graphics card** (display driver only, nothing else to install — |
| > about 10× faster answers and X-rays than 3.1 on an RTX laptop), **nearly twice as fast on a |
| > Windows processor**, Qwen3.5 reading requests 8–10× faster, instant follow-up questions, |
| > and **MedSearch that understands everyday words and abbreviations** ("water infection", |
| > "AF", "piles"). MedSearch knowledge packs (CloudKites' summaries of eTG, AMH and Australian |
| > guidelines) are included. The previous edition, |
| > [MedPod GP 2.0](#-previous-edition-medpod-gp-20-windows) (Windows), is still available. |
|
|
| > 📘 **New to MedPod GP? Start with the [User Guide (PDF)](https://huggingface.co/cloudkites/medpodgp/resolve/main/USER_GUIDE.pdf).** |
| > Installing on Mac, Windows or Linux, the first start, and every feature step by |
| > step, with screenshots. |
|
|
| --- |
|
|
| ## 🎬 Demonstration — the same app on Mac, Windows and Linux |
|
|
| MedPod GP 3.1 running its own live self-test, paced so each step can be read: a new |
| practice account, a 75-second dictation transcribed, **Summarise**, a MedText question |
| and follow-up, a chest X-ray report, a **MedSearch** question answered from the knowledge |
| packs with citations, the draft and final note, saving, exporting and locking. Real time, |
| not sped up; the made-up consultation of the User Guide; no sound. |
|
|
| **🍎 Mac** (MacBook Pro M5 Pro; MedGemma on the Apple GPU) |
|
|
| <video controls width="100%" src="https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-macos.mp4"></video> |
|
|
| **🪟 Windows** (ThinkPad, Core i7-13800H laptop; on the processor) |
|
|
| <video controls width="100%" src="https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-windows.mp4"></video> |
|
|
| **🐧 Linux** (Minisforum MS-A2, Ryzen 9 9955HX, KDE Plasma; on the processor) |
|
|
| <video controls width="100%" src="https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-linux.mp4"></video> |
|
|
| Video files: [Mac](https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-macos.mp4) · |
| [Windows](https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-windows.mp4) · |
| [Linux](https://huggingface.co/cloudkites/medpodgp/resolve/main/demo/medpodgp-3.1-linux.mp4). |
| The Windows and Linux videos show the window as the desktop draws it; the Mac video is |
| the window's own content, recorded by the app. |
|
|
| --- |
|
|
| ## ⬇️ Download MedPod GP 3.2 — pick your computer |
|
|
| The same app, with the same interface, on all three systems — native on each, with |
| **no runtime dependencies** (nothing else to install). |
|
|
| | Your computer | Download | Install | |
| |---|---|---| |
| | 🍎 **Mac** — Apple silicon (M1 or later), macOS 13+ | [**medpodgp-3.2.1-macos-arm64.dmg**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medpodgp-3.2.1-macos-arm64.dmg) | Open it and drag **MedPod GP** to **Applications** | |
| | 🪟 **Windows 10/11** (64-bit) | [**medpodgp-3.2.1-windows-x64.zip**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medpodgp-3.2.1-windows-x64.zip) | Unzip anywhere and double-click **`medpodgp.exe`** — no installer | |
| | 🐧 **Linux** (x86-64, glibc 2.31+) | [**medpodgp-3.2.1-linux-x64.tar.gz**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medpodgp-3.2.1-linux-x64.tar.gz) | Extract, then `./install.sh` (or run `./medpodgp` in place) | |
|
|
| Every file has a matching **`.sha256`**, and |
| [**`SHA256SUMS`**](https://huggingface.co/cloudkites/medpodgp/resolve/main/SHA256SUMS) |
| lists them all. |
|
|
| **First launch.** |
|
|
| - **Mac:** MedPod GP is signed with CloudKites' Apple Developer ID and **notarized by |
| Apple** (the ticket is stapled, so it opens offline too). The first time, macOS asks |
| whether to open an app downloaded from the internet — click **Open**. |
| - **Windows:** `medpodgp.exe` is code-signed by **CLOUDKITES PTY LTD**. SmartScreen |
| may still warn about a brand-new release — click **More info**, check the |
| publisher, then **Run anyway**. |
| - **Linux:** `./install.sh` installs for you only (no sudo): `~/.local/bin/medpodgp`, |
| an applications-menu entry and an icon. `./install.sh --remove` uninstalls. |
| Install **zenity** or **kdialog** for Open/Save dialogs. |
|
|
| On first start MedPod GP asks for the **practice owner's account** (local, on this |
| computer), then shows **Download the AI models**: it downloads and verifies |
| (SHA-256) the models you choose, with pause and resume. After that, everything runs |
| offline. |
|
|
| 📄 **Documents:** |
| [User Guide](https://huggingface.co/cloudkites/medpodgp/resolve/main/USER_GUIDE.pdf) · |
| [Product Information](https://huggingface.co/cloudkites/medpodgp/resolve/main/PRODUCT_INFORMATION.pdf) · |
| [Download & Installation Guide](https://huggingface.co/cloudkites/medpodgp/resolve/main/DOWNLOAD_INSTALL_GUIDE.pdf) · |
| [License](https://huggingface.co/cloudkites/medpodgp/resolve/main/LICENSE.md) |
|
|
| --- |
|
|
| ## 🧠 The AI models |
|
|
| Downloaded inside the app on first use — or download them here and put them in |
| the models folder for a computer that must stay offline (the app finds them; see |
| the Download & Installation Guide). |
|
|
| | Model pack (`.nbq`) | Powers | Size | Licence | |
| |---|---|---|---| |
| | [**medgemma-1.5-4b-q4km.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medgemma-1.5-4b-q4km.nbq) | **MedGemma 1.5 4B** — medical text **and images** (MedVision); the recommended model | 3.4 GB | Health AI Developer Foundations Terms of Use | |
| | [**qwen3.5-4b-q4km.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3.5-4b-q4km.nbq) | **Qwen3.5 4B** — general text, a second opinion | 2.8 GB | Apache-2.0 | |
| | [**qwen3.5-2b-q4km.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3.5-2b-q4km.nbq) | **Qwen3.5 2B** — light and fast, for 8 GB computers | 1.3 GB | Apache-2.0 | |
| | [**qwen3-asr-0.6b.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3-asr-0.6b.nbq) | **Qwen3-ASR 0.6B** — speech recognition for MedAudio | 1.9 GB | Apache-2.0 | |
|
|
| Models folder: `~/Library/Application Support/medpodgp/models` (Mac), |
| `%LOCALAPPDATA%\medpodgp\models` (Windows), `~/.local/share/medpodgp/models` (Linux). |
| Each pack is one self-contained file: the model's weights (4-bit for the language |
| models) with its tokenizer and, for MedGemma, its image encoder. |
|
|
| --- |
|
|
| ## ✨ Features |
|
|
| - **🎤 MedAudio** — record the consultation (or open a recording) and transcribe it |
| on the computer with Qwen3-ASR (10 languages or automatic detection); long |
| consultations are transcribed in pieces as they go. Then **Summarise**, |
| **Diagnosis** (differentials with reasoning), **Advise** and **Questions**. |
| - **💬 MedText** — clinical questions in plain words, with fast follow-ups. |
| - **🩻 MedVision** — X-ray, CT, MRI, mammogram, fundoscopy, skin, histopathology, |
| endoscopy, ultrasound and ECG, each with its own structured report and ten body |
| regions; follow-up questions about the image (MedGemma 1.5). |
| - **🔎 MedSearch** — search the three knowledge packs that come with the app |
| (CloudKites' summaries of Therapeutic Guidelines, the Australian Medicines Handbook |
| and Australian clinical guidelines); **Ask AI** answers from the best matches, |
| every sentence cited and each citation checked against its source. |
| - **📝 The medical note** — add any answer (or part of one) under the section you |
| choose; **Generate final note** writes a structured clinical note; copy it, save |
| it, or export **PDF, Word, Markdown or text**. |
| - **🔐 For the whole practice** — local accounts (owner, administrators, |
| clinicians), private notes, automatic lock, an audit log of who did what. |
| - **💻 Native everywhere** — one app on macOS, Windows and Linux; the Mac's GPU, an |
| NVIDIA graphics card on Windows (driver only), every core of a processor otherwise; |
| light and dark themes. |
|
|
| ## 🎉 What's new in 3.2.1 |
|
|
| - **Speech on the NVIDIA card from start to finish** — in 3.2 the recognizer heard the |
| recording on the graphics card but wrote the text on the processor; now both run on the |
| card. On an RTX 2000 Ada laptop 75 s of speech is ready in **4.5 s instead of 11 s**, |
| loading included, and the recognizer now **loads while you record**, so the text follows |
| the stop button by about 2 s. |
| - **Far less memory on Windows** — the recognizer and the image encoder each held their |
| weights twice (in the computer's memory and on the card, which Windows charges to the |
| computer too) and kept memory they had finished with. A recording on the card now needs |
| **3.5 GB** of Windows' committed memory (9.6 GB in 3.2.0), an X-ray **4.9 GB** (7.5), and |
| speech on the processor 4.1 GB (7.6), so the language model steps aside far less often. |
| - **Lighter everywhere** — the recognizer loads in about half the time, and every system |
| needs less memory: 75 s of speech 7.5 GB on the Mac (11.2 in 3.2.0) and 3.7 GB on Linux |
| (6.0); an X-ray 5.6 GB on the Mac (7.5). |
|
|
| ## 🎉 What's new in 3.2 |
|
|
| - **Windows runs the AI on the NVIDIA graphics card** — many clinical laptops have one. |
| Nothing extra to install: the display driver is enough (the GPU programs are built into |
| MedPod GP, which stays one signed `.exe`). On an RTX 2000 Ada laptop: answers at |
| **71 tokens/s instead of 7.5**, a consultation transcript read in **0.4 s instead of 11 s**, |
| a chest X-ray read in **under 5 s instead of 57 s**. Without such a card the processor takes over |
| by itself. |
| - **Nearly twice as fast on a Windows processor** — Windows 11's efficiency mode was slowing |
| MedPod GP down; it now asks to run at full speed: 13 tokens/s on the laptop's processor |
| (7.5), an X-ray in 32 s (57), 75 s of speech in 14 s (23). |
| - **Qwen3.5 reads requests 8–10× faster on processors** (in one batch instead of a token at |
| a time), and **follow-up questions start in a fraction of a second** on every computer. |
| - **Memory leaks fixed** — each image used to leave 1.7 GB behind, and each transcription |
| on the processor about 300 MB (some thirty recordings ran a Windows laptop out of memory). |
| - **Memory planned on Windows** — Windows counts the graphics card's memory against the |
| computer's own; MedPod GP checks what is left before an image or a recording and, when |
| it is short, sets the language model aside for a moment instead of failing. |
| - **MedSearch understands how GPs write** — "water infection", "piles", "AF", "blood |
| thinner", "pink eye" and 100+ other everyday phrases and abbreviations find the pages |
| that say "urinary tract infection", "haemorrhoids", "atrial fibrillation", |
| "anticoagulant", "conjunctivitis"; US spellings find the Australian ones; "cervical" |
| means the cervix with *screening* and the neck with *radiculopathy*. On held-out |
| questions the first result is relevant **95% of the time (76% in 3.1.1)**. |
|
|
| ## 🎉 What's new in 3.1.1 |
|
|
| - **Sharper MedSearch.** Measured on 26 held-out GP questions: a relevant source in the top |
| four for all of them (92% in 3.1.0), and **83% of the sources Ask AI reads are relevant** |
| (38% in 3.1.0) — it now reads only the sources that score close to the best one (about 2 |
| instead of 4), so answers stay on topic. Sources written for another setting (pregnancy, |
| children, dentistry, palliative care, poisoning, …) and a drug's eye, rectal or |
| combination forms no longer push out the page the question is about. |
|
|
| ## 🎉 What's new in 3.1 |
|
|
| - **MedSearch knowledge packs included** — 2,617 sources in 18,096 sections, ready the |
| first time the app opens: summaries written by CloudKites of eTG, the AMH and |
| Australian clinical guidelines (the publishers' own text is not included; every |
| section names and links its source). |
| - **Speech recognition about seven times faster on Windows and Linux** — 75 s of |
| speech in about 23 s on a Core i7 laptop (2 min 39 s in 3.0) and 6 s on a Ryzen 9; |
| faster than the recording plays on every machine measured. |
| - **Every Ask AI sentence cited** — where the model leaves a sentence uncited, the app |
| adds the source that carries it; better source ranking across the packs. |
|
|
| ## 🎉 What's new in 3.0 |
|
|
| - **Mac, Windows and Linux** — a native window on each, no Qt, no installer on |
| Windows, a per-user install on Linux. |
| - **numbat SDK** — CloudKites' on-device engine: MedGemma on the Apple GPU at about |
| 90 tokens/s; on Windows and Linux, batched prompt reading and a thread count tuned |
| to the processor (performance cores on Intel hybrid chips). |
| - **One verified file per model**, downloaded in the app with pause/resume. |
| - **Practice accounts** with roles, private notes, auto-lock and an audit log, in |
| place of 2.0's built-in administrator. |
| - **MedSearch Ask AI** with a citation check, and **guarded answers** that end |
| cleanly instead of repeating. |
| - **Exports** to PDF and Word. |
|
|
| ## ⚡ Measured speed |
|
|
| MedGemma 1.5 4B summarising a consultation transcript (~550 tokens in, 300 out), a |
| chest X-ray report, and transcription of 75 s of speech: |
|
|
| | | Mac (M5 Pro, GPU) | Linux (Ryzen 9 9955HX, CPU) | Windows laptop (i7-13800H, CPU) | Windows laptop (RTX 2000 Ada) | |
| |---|---|---|---|---| |
| | Writing speed | 87 tok/s | 19 tok/s | 13 tok/s | 71 tok/s | |
| | Wait before the summary starts | 0.2 s | 2.2 s | 7.3 s | 0.4 s | |
| | Wait before a follow-up answer | 0.06 s | 0.1 s | 0.4 s | 0.2 s | |
| | X-ray: reading the image | 1.1 s | 10 s | 32 s | 4.6 s | |
| | 75 s of speech transcribed in | 1.4 s | 5.2 s | 13 s | 4.5 s | |
|
|
| The Windows GPU column is the display driver alone, as a GP's PC has it; its X-ray time |
| includes setting the language model aside for a moment, which MedPod GP does when Windows |
| is short of memory. Speech times include loading the recognizer, which in the app happens |
| during the recording (on an NVIDIA card the text is ready about 2 s after stop). People read about 5–7 tokens a second, so answers keep pace with reading on every |
| machine measured, and every one transcribes faster than the recording plays. The |
| Product Information PDF has the full charts. |
|
|
| ## 💻 System requirements |
|
|
| | | Mac | Windows | Linux | |
| |---|---|---|---| |
| | System | macOS 13+ | Windows 10 or 11, 64-bit | glibc 2.31+ (Ubuntu 20.04+, Debian 11+, Fedora 32+) | |
| | Processor | Apple M1 or later | AVX2 (Intel 2013+, AMD 2015+) | AVX2 | |
| | Graphics | (the Mac's own GPU) | optional: an NVIDIA GeForce RTX 20 series or newer / RTX workstation card (6 GB for images), display driver only | — | |
| | Memory | 16 GB (8 GB with Qwen3.5 2B) | 16 GB | 16 GB | |
| | Disk | 8 GB for MedGemma + speech | 8 GB | 8 GB | |
|
|
| ## 📚 MedSearch knowledge packs |
|
|
| Three packs come inside every download — nothing to install: |
|
|
| | Pack | Summarises | Sources | Sections | |
| |---|---|---|---| |
| | eTG summaries | Therapeutic Guidelines (eTG complete) | 1,305 | 10,319 | |
| | AMH summaries | Australian Medicines Handbook | 1,180 | 3,439 | |
| | Guideline summaries | Australian clinical guidelines (RANZCOG, RCH Melbourne, cervical screening, blood management, pregnancy care, …) | 132 | 4,338 | |
|
|
| Each section is a summary written by CloudKites, not the publisher's text, and names |
| and links the guideline it summarises. A pack file (`.sqlite`) placed in the data |
| folder's `packs` folder replaces the bundled one of the same name. |
|
|
| --- |
|
|
| ## 🗂 Previous edition: MedPod GP 2.0 (Windows) |
|
|
| MedPod GP 2.0 — the Windows edition built on Qt, with CPU and NVIDIA CUDA |
| installers — remains available for existing installations. Version 3 replaces it |
| on every platform; it keeps its own database and model packs. |
|
|
| | Build | Installer | |
| |---|---| |
| | CPU (AVX2) | [MedPodGPInstaller-2.0.0-cpu.exe](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu.exe) | |
| | CPU, Intel 13th/14th gen (AVX-VNNI) | [MedPodGPInstaller-2.0.0-cpu-raptorlake.exe](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu-raptorlake.exe) | |
| | CPU, AVX-512 | [MedPodGPInstaller-2.0.0-cpu-avx512.exe](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu-avx512.exe) | |
| | NVIDIA GPU (CUDA) | [MedPodGPInstaller-2.0.0-cuda.exe](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cuda.exe) | |
|
|
| 2.0's model bundles (`*-cpu.nbq`, `*-cuda.nbq`) are on this page too, and its |
| documents are in [`v2/`](https://huggingface.co/cloudkites/medpodgp/tree/main/v2): |
| [User Guide](https://huggingface.co/cloudkites/medpodgp/resolve/main/v2/USER_GUIDE.pdf) · |
| [Product Information](https://huggingface.co/cloudkites/medpodgp/resolve/main/v2/PRODUCT_INFORMATION.pdf). |
|
|
| --- |
|
|
| ## ⚕️ Medical disclaimer & intended use |
|
|
| **MedPod GP is a clinical decision-support and documentation aid intended for use |
| by qualified healthcare professionals. It is NOT a medical device and has not been |
| cleared or approved by any medical-device regulator (e.g. TGA, FDA, CE/MDR).** |
|
|
| - It does **not** diagnose, treat, cure, or prevent any disease or condition. |
| - It does **not** replace the clinical judgement of a qualified clinician. |
| - **All AI-generated output — transcripts, answers, image reports, note drafts, and |
| search answers — must be reviewed, corrected, and verified by a qualified |
| clinician before any clinical use or entry into a patient record.** |
| - The responsible clinician remains accountable for all clinical decisions and |
| documentation. |
| - Do not rely on MedPod GP in emergencies or as a sole source of clinical |
| information. |
|
|
| --- |
|
|
| ## 📜 License |
|
|
| MedPod GP is **proprietary software distributed free of charge** © 2026 CloudKites |
| Pty Ltd (CloudKites AI Lab), ABN 91 633 195 874, in compiled (binary) form for |
| macOS, Windows and Linux. |
|
|
| - ✅ **Free to use** — on any number of devices. |
| - ✅ **Free to redistribute** — the **complete, unmodified** official package, free |
| of charge, with all notices intact. |
| - ❌ **No selling / no paid bundling**, no modification, no reverse-engineering. |
| - The source code of MedPod GP and of its **Numbat** ML toolkit is confidential and |
| is not distributed. |
|
|
| Full terms: [`LICENSE.md`](https://huggingface.co/cloudkites/medpodgp/resolve/main/LICENSE.md). |
| "MedPod GP", "Numbat", and "CloudKites" are trademarks of CloudKites Pty Ltd. |
|
|
| The AI models remain under their own licences — |
| [`THIRD_PARTY_LICENSES.md`](https://huggingface.co/cloudkites/medpodgp/resolve/main/THIRD_PARTY_LICENSES.md). |
| MedGemma is provided under and subject to the **Health AI Developer Foundations |
| Terms of Use** (https://developers.google.com/health-ai-developer-foundations/terms), |
| including its use restrictions; the Qwen models are under Apache-2.0. |
|
|
| --- |
|
|
| ## ⚠️ AI output notice |
|
|
| MedPod GP uses generative AI models. Output may be inaccurate, incomplete, or |
| fabricated ("hallucinated") even when it appears confident, and is **not** a |
| substitute for professional clinical judgement. **A qualified clinician must verify |
| every output before use.** |
|
|
| --- |
|
|
| Contact: **contact@cloudkites.com** · **cloudkites.com** |
|
|
