Portable AI Studio: Run Offline AI Chat, Images and Voice From a USB Drive

Source: Author

Imagine a USB drive that opens a private AI workspace in your browser with one double-click, even with the Wi-Fi off. It has a chat assistant, an image generator, a transcriber and a voice that reads text aloud. No sign-up, no server, and your history goes wherever the drive goes.

That is the idea behind Portable AI Studio, a free, MIT-licensed open-source project. After a close look at its code and documentation, most of the promise holds up. The exceptions are worth knowing before you buy a drive, because they decide whether your studio feels quick or crawls.

Four AI tools in one browser app

The app opens at localhost:1420 and wraps four well-known open-source engines:

• Chat runs GGUF language models through llama.cpp. You can attach PDFs, text or code files, and images once a vision model is loaded.
• Images runs Stable Diffusion 1.5 and SDXL checkpoints through stable-diffusion.cpp, saving each result to a gallery with its prompt and settings.
• Transcribe converts speech from the microphone or an audio file with whisper.cpp.
• Voice reads text aloud with Kokoro-82M, and you can download the audio.

It uses the fastest path your hardware offers: CUDA on Nvidia, Vulkan on AMD and Intel Arc, Metal on Apple Silicon, and OpenVINO on the NPU of Intel Core Ultra chips for images. Machines without a usable GPU still run it on the CPU, just more slowly.

The drive holds the app, the models and your history, while the computer's GPU, RAM and CPU do the work

Setting up the drive

1. Get at least 32 GB, USB 3.0 or faster. Each model switch reads from the drive, so a portable SSD helps most.
2. Format it as exFAT so Windows, macOS and Linux can all read and write it. This erases the drive.
3. Download the ZIP from the project's GitHub page and extract it onto the drive.
4. Run windows.bat, mac.sh or linux.sh. The first run fetches a portable Node.js runtime and the engine builds.
5. Open Models, look for the "Fits your PC" (or "Fits your Mac") dot, and download straight to the drive.

Space goes quickly. SDXL models in the app's library are about 7 GB each and most SD 1.5 models about 2 GB, so pick 64 GB or more if you want several image models.

Which computers it runs on

Windows needs 64-bit Windows 10 or 11. Macs need Apple Silicon (M1 or newer), and the Transcribe tool there installs its engine through Homebrew. On Linux, the prebuilt engines expect glibc 2.38 or later, so think Ubuntu 24.04 or something equally recent. Your models and history travel with the drive, but each operating system downloads its own runtime on its first launch.

Model tiers come mostly from graphics memory. A PC with about 6 GB of VRAM, or 16 GB of system RAM, lands in Mid, and about 12 GB of VRAM reaches High. On Apple Silicon, 8 GB of unified memory means Mid and 16 GB or more puts you in the top tier.

Starter chat, image, transcription and voice models for Low, Mid and High tier machines

The catches

• Setup needs the internet. Runtimes and models download first, and the voice model fetches its files the first time you press Speak.
• Web search uses the internet. The Chat toggle sends a query to DuckDuckGo, so leave it off when privacy is the point.
• There is no login. The setting UAIS_HOST=0.0.0.0 lets anyone on the network use the studio, so never switch it on over café or hotel Wi-Fi.
• exFAT has no encryption. Losing the drive exposes your files unless the drive has hardware encryption.
• Choose downloads carefully. Stick to known publishers, and pick .safetensors files over older .ckpt ones.

Read the full Portable AI Studio guide on Medium →

Mac and Linux launch steps, extra options for AMD and Intel hardware, and where a pocket AI studio earns its place.