| ---
|
| license: other
|
| license_name: medpodgp-eula
|
| license_link: LICENSE.md
|
| language:
|
| - en
|
| tags:
|
| - medical
|
| - clinical
|
| - healthcare
|
| - on-device
|
| - local-ai
|
| - speech-to-text
|
| - desktop-app
|
| pretty_name: MedPod GP — On-Device Medical AI Assistant
|
| ---
|
|
|
| # MedPod GP — On-Device Medical AI Assistant
|
|
|
| **MedPod GP** is a **desktop medical AI assistant** for clinicians and general
|
| practitioners that runs **entirely on your own computer**. Dictate and transcribe
|
| clinical encounters, chat with a medical language model, analyse medical images,
|
| search trusted reference packs, and draft structured EMR-ready notes — all
|
| **on-device**, with **no cloud account, no telemetry, and no network calls for
|
| inference**.
|
|
|
| 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
|
| > installer** for free (no selling, no paid bundling). See **License** below.
|
|
|
| > ⚕️ **Clinical decision-support / documentation aid — not a medical device.**
|
| > MedPod GP does not diagnose, treat, or replace professional medical judgement.
|
| > All AI output must be reviewed and verified by a qualified clinician. See the
|
| > **Medical disclaimer & intended use** section below.
|
|
|
| ---
|
|
|
| ## ⬇️ Installation (Windows) — v2.0
|
|
|
| **New in 2.0: one small app installer, models on first run.** You install a single
|
| **code-signed app** (~80 MB). The first time you launch it, MedPod GP downloads its
|
| AI models — now packaged as **single self-contained `.nbq` files** — straight into
|
| your local models folder, with a progress screen. After that, **nothing is
|
| downloaded again and everything runs on-device.**
|
|
|
| > 🔒 **Offline & privacy — unchanged.** MedPod GP still runs **100% on your
|
| > computer**: no cloud account, no telemetry, and **no network calls for any AI
|
| > inference**. The **only** time it uses the internet is to **download the model
|
| > files once** (or when you add a feature). Prefer to stay fully offline? **Download
|
| > the `.nbq` files manually** (table below) and drop them into your models folder —
|
| > the app finds them and never touches the network. See *“Fully offline / manual
|
| > install”* below.
|
|
|
| ### Step 1 — Install the MedPod GP app
|
|
|
| Download **one** build and run the installer (code-signed, Microsoft Azure Trusted
|
| Signing):
|
|
|
| | Build | For | Installer |
|
| |---|---|---|
|
| | **CPU** | Any 64-bit Windows PC with **AVX2** (Intel Haswell 2013+ / AMD 2015+) | [**MedPodGPInstaller-2.0.0-cpu.exe**](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu.exe) |
|
|
|
| **Optimized CPU builds (advanced).** Faster prompt processing / image analysis on
|
| newer CPUs. **Each requires the matching CPU feature and won't run otherwise** — if
|
| unsure, use the standard **CPU** build (it runs everywhere).
|
|
|
| | Optimized build | Best for | Requires |
|
| |---|---|---|
|
| | [**MedPodGPInstaller-2.0.0-cpu-raptorlake.exe**](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu-raptorlake.exe) | Intel 13th/14th-gen Core (i5/i7/i9) — adds **AVX-VNNI** | Intel 12th-gen+ Core (AVX-VNNI) |
|
| | [**MedPodGPInstaller-2.0.0-cpu-avx512.exe**](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cpu-avx512.exe) | AMD Ryzen 7000/9000 (Zen 4/5) and AVX-512 Intel | Any CPU with AVX-512 |
|
|
|
| **GPU (CUDA) build.** Have a supported NVIDIA card? Use the CUDA build for the
|
| fastest transcription, chat, and image analysis. It **bundles the CUDA + cuDNN
|
| runtime** — nothing extra to install.
|
|
|
| | GPU build | Best for | Requires |
|
| |---|---|---|
|
| | [**MedPodGPInstaller-2.0.0-cuda.exe**](https://huggingface.co/cloudkites/medpodgp/resolve/main/MedPodGPInstaller-2.0.0-cuda.exe) | Any PC with an NVIDIA GPU (**Compute Capability ≥ 7.0**) | NVIDIA GPU + recent driver |
|
|
|
| ### Step 2 — First launch downloads the models
|
|
|
| Launch **MedPod GP**. If the model files aren't present yet, it shows a **first-time
|
| setup** screen that downloads the bundles it needs and shows progress. When it
|
| finishes, sign in and start working — **no further downloads, ever.** A download that
|
| gets interrupted **resumes** where it left off when you retry.
|
|
|
| ### Fully offline / manual install
|
|
|
| You can skip the in-app download entirely: grab the `.nbq` files below and place them
|
| in your models folder — `%LOCALAPPDATA%\medpodgp\models` (e.g.
|
| `C:\Users\<you>\AppData\Local\medpodgp\models`). The app detects them on launch and
|
| runs without any network access.
|
|
|
| **Match the bundles to your build.** Two of the models (text, summariser) are
|
| **device-neutral** — the same file works on the CPU and CUDA builds. The other two
|
| (vision, ASR) ship **device-specific** weights, so a CPU build uses the `-cpu`
|
| bundle and a CUDA build uses the `-cuda` bundle. Download the **two shared** files
|
| plus the **two** that match your build.
|
|
|
| **Shared by both builds:**
|
|
|
| | Model bundle (`.nbq`) | Powers | Required? | Size |
|
| |---|---|---|---|
|
| | [**medgemma-1.5-4b-text-q4km-cpu.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medgemma-1.5-4b-text-q4km-cpu.nbq) | **MedText** chat + AI note generation — MedGemma 1.5 (text) | **Required** | ~2.5 GB |
|
| | [**qwen3.5-0.8b-summariser-q4km-cpu.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3.5-0.8b-summariser-q4km-cpu.nbq) | Auto-summarising long inputs before chat — Qwen3.5 0.8B | Recommended | ~0.5 GB |
|
|
|
| **CPU build only:**
|
|
|
| | Model bundle (`.nbq`) | Powers | Required? | Size |
|
| |---|---|---|---|
|
| | [**medgemma-1.5-4b-vision-cpu.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medgemma-1.5-4b-vision-cpu.nbq) | **MedVision** image analysis — MedGemma 1.5 (vision) | Required for MedVision | ~0.85 GB |
|
| | [**qwen3-asr-0.6b-int8-cpu.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3-asr-0.6b-int8-cpu.nbq) | **MedAudio** voice transcription — Qwen3-ASR | Required for MedAudio | ~1.1 GB |
|
|
|
| **CUDA (GPU) build only:**
|
|
|
| | Model bundle (`.nbq`) | Powers | Required? | Size |
|
| |---|---|---|---|
|
| | [**medgemma-1.5-4b-vision-cuda.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/medgemma-1.5-4b-vision-cuda.nbq) | **MedVision** image analysis — MedGemma 1.5 (vision, F16) | Required for MedVision | ~0.85 GB |
|
| | [**qwen3-asr-0.6b-f16-cuda.nbq**](https://huggingface.co/cloudkites/medpodgp/resolve/main/qwen3-asr-0.6b-f16-cuda.nbq) | **MedAudio** voice transcription — Qwen3-ASR (F16) | Required for MedAudio | ~1.5 GB |
|
|
|
| Each `.nbq` is a single self-contained, integrity-checked file holding the model's
|
| (unchanged, same-quality) weights **plus its tokenizer/config** — so the app loads
|
| one file and is ready. The first-run downloader automatically fetches the right set
|
| for your build; the table above is only for manual/offline placement. **MedSearch**
|
| reference search needs no model.
|
|
|
| ### Signing in & accounts
|
|
|
| MedPod GP has **local user accounts** (stored only on your PC — no online account, no
|
| telemetry). There are **administrators** (can use every feature *and* manage accounts)
|
| and **standard users** (every clinical feature, no account management).
|
|
|
| **First sign-in — built-in administrator:**
|
|
|
| | Username | Password | Role |
|
| |---|---|---|
|
| | `nikonsugar` | `cloudkitesailab` | Administrator |
|
|
|
| On the login screen, enter these, **tick “Login as administrator,”** and click **Sign
|
| in**. 🔐 **Change this right away** — it's the same on every install, so it isn't
|
| secret. Create your own admin with a strong password and use that instead.
|
|
|
| - **Create a user:** signed in as an admin, open the **administrator dashboard →
|
| Accounts**, click **Add User**, set a username + password (+ optional display name),
|
| **Save**. (Use **Add Admin** to grant another person admin rights.)
|
| - **Log out / switch user:** click **Logout** in the top bar; sign in as the other
|
| user — leave **“Login as administrator” unticked** for a standard account.
|
|
|
| Full step-by-step instructions are in the **User Guide** below.
|
|
|
| **Documentation:**
|
| [User Guide (PDF)](https://huggingface.co/cloudkites/medpodgp/resolve/main/USER_GUIDE.pdf) ·
|
| [Product Information (PDF)](https://huggingface.co/cloudkites/medpodgp/resolve/main/PRODUCT_INFORMATION.pdf) ·
|
| [License](https://huggingface.co/cloudkites/medpodgp/resolve/main/LICENSE.md)
|
|
|
| ---
|
|
|
| ## ✨ Features
|
|
|
| MedPod GP combines five clinical capabilities in one private, offline workspace:
|
|
|
| - **🎤 MedAudio — clinical transcription.** Record a consultation live (with a
|
| real-time audio spectrogram) and transcribe it on-device with **Qwen3-ASR
|
| 0.6B**, then run clinical actions on the transcript.
|
| - **💬 MedText — medical chat.** Ask clinical questions of a hot-swappable medical
|
| language model: **MedGemma 1.5 4B**, **Qwen3.5 4B**, or **Qwen3.5 2B**. Long
|
| inputs are auto-summarised before the model reads them.
|
| - **🩻 MedVision — medical image analysis.** Upload a medical image and analyse it
|
| with the **MedGemma 1.5** vision pipeline, with structured prompts for around
|
| **10 imaging types**.
|
| - **🔎 MedSearch — reference search.** Fast full-text search over bundled medical
|
| reference packs (eTG, AMH, foundation) — answers in milliseconds, fully offline.
|
| - **📝 Medical Note Drafting.** A cross-tab panel that captures content from any
|
| tab into a structured clinical note and generates an **EMR-formatted final
|
| note**, with **Markdown / PDF export**.
|
|
|
| - **100% private & offline.** Every model runs on your CPU or NVIDIA GPU; nothing
|
| leaves your device. No account, no telemetry.
|
|
|
| ---
|
|
|
| ## 💻 Which build for my computer?
|
|
|
| | Your PC | Recommended build | Notes |
|
| |---|---|---|
|
| | Has an NVIDIA GPU (Compute Capability ≥ 7.0) | **GPU (CUDA)** | Fastest transcription, chat, and image analysis. Bundles CUDA/cuDNN — nothing extra to install |
|
| | CPU-only, modern processor with **AVX2** (≈ 2014 onward), 16 GB+ RAM | **CPU** | Runs every feature; large models and image analysis are slower |
|
|
|
| If you are unsure whether your processor supports AVX2, the **CPU** build runs on
|
| essentially every Windows PC built since about 2014. Choose the **GPU (CUDA)**
|
| build whenever you have a supported NVIDIA graphics card.
|
|
|
| ---
|
|
|
| ## ⚡ Expected performance (rough guide)
|
|
|
| AI speed is described by two numbers, both in **tokens per second (tok/s)**:
|
|
|
| - **Decode** — how fast a reply or transcript is *generated*. This is limited by
|
| your **RAM speed** (or GPU memory), **not** by the CPU model — so faster RAM, and
|
| **dual-channel** (two sticks) instead of single-channel, matters most.
|
| - **Prefill / prompt** — how fast a prompt (or image) is *read in* before the reply
|
| starts. This is limited by **CPU compute**, so more cores and the optimized
|
| AVX-VNNI / AVX-512 builds help here.
|
|
|
| Rough estimates for the two main models — **MedGemma 1.5 4B** (chat / vision) and
|
| **Qwen3-ASR 0.6B** (transcription), both 4-bit. Treat as ballpark (±~30 %); MedPod
|
| GP shows live tok/s so you can see your real numbers.
|
|
|
| | Your hardware (typical) | MedGemma decode | MedGemma prefill | ASR decode | What it feels like |
|
| |---|---|---|---|---|
|
| | **CPU · AVX2** — Intel ~2014–2021 / AMD Zen 1–3, DDR4-3200 dual | ~10–14 tok/s | ~60–150 tok/s | ~50–90 tok/s | Chat a touch slower than reading; transcription faster than real-time |
|
| | **CPU · Raptor Lake** — Intel 13/14th-gen, DDR5-5600 dual | ~15–22 tok/s | ~120–280 tok/s | ~80–130 tok/s | Comfortable reading-pace chat; snappy prompts & images |
|
| | **CPU · AVX-512** — AMD Ryzen 7000/9000 desktop, DDR5-6000 dual | ~16–24 tok/s | ~150–350 tok/s | ~90–150 tok/s | Fastest CPU prompt & image analysis |
|
| | **CPU · AVX-512 server** — Xeon / EPYC, 8–12-channel DDR5 | ~40–90 tok/s | very high | ~150+ tok/s | Workstation-class throughput |
|
| | **GPU (CUDA)** — RTX 3060–4070-class | ~50–110 tok/s | very high | near-instant | Smooth, real-time everything |
|
|
|
| **Notes**
|
| - **Single-channel RAM ≈ half the decode speed** of the figures above — use two
|
| matched sticks for dual-channel.
|
| - The optimized **Raptor Lake / AVX-512** builds mainly speed up **prefill and
|
| MedVision image analysis** (compute-bound); decode is RAM-bound, so the gain there
|
| over the standard CPU build is small.
|
| - **16 GB RAM** comfortably runs chat + transcription. MedVision, very long chats,
|
| or heavy multitasking need more — if RAM runs out the PC pages to disk and
|
| everything slows sharply (32 GB is safer for those).
|
|
|
| ---
|
|
|
| ## ⚕️ 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, chat responses, image findings, note
|
| drafts, and search results — 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. It is provided in compiled
|
| (binary) form for Windows.
|
|
|
| - ✅ **Free to use** — for qualified clinicians and healthcare settings, on any
|
| number of devices.
|
| - ✅ **Free to redistribute** — you may share the **complete, unmodified** official
|
| installer, free of charge, with all notices intact.
|
| - ❌ **No selling / no paid bundling**, no modification, no reverse-engineering.
|
| - The source code of MedPod GP and 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.
|
|
|
| Bundled AI models remain under their own licenses — see
|
| [`THIRD_PARTY_LICENSES.md`](https://huggingface.co/cloudkites/medpodgp/resolve/main/THIRD_PARTY_LICENSES.md).
|
| Use of the MedGemma / Gemma models is additionally subject to the **Gemma Terms of
|
| Use** and Prohibited Use Policy.
|
|
|
| > **Note on models.** The AI models are provided as separate **code-signed Windows
|
| > model installers** on this page (see **Installation → Step 2**). Each bundled
|
| > model remains under its own license — use of the MedGemma / Gemma models is
|
| > subject to the **Gemma Terms of Use** and Prohibited Use Policy.
|
|
|
| ---
|
|
|
| ## ⚠️ 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**
|
|
|
