opnlocal 0.2.2

0.2.2

Fix: the phone apps now show opnlocal's icon. The Android and iPhone builds shipped with the app framework's placeholder icon. Nothing else changed since 0.2.1.

Install

Download the file for your device from the assets below, or use the direct links on opnlocal.pages.dev. Windows shows a SmartScreen warning because the app is not yet code-signed: More info → Run anyway. On a Mac, right-click the app → Open the first time.

opnlocal 0.2.1

0.2.1

Fix: attaching photos and PDFs on Android. Files picked from the gallery or Downloads came without a file name, so the app answered "this file type isn't supported". The type is now read from the file itself, on every platform.

Everything from 0.2.0 is included: images (text read on the device, or the picture itself for models with image support), documents of any length, four more models.

Install

Download the file for your device from the assets below, or use the direct links in the README. Windows shows a SmartScreen warning because the app is not yet code-signed: More info → Run anyway. On a Mac, right-click the app → Open the first time.

opnlocal 0.2.0

What's new in 0.2.0

Images. Attach a photo or screenshot. Its text is read on your device (no upload), and models that can look at pictures do so after an extra download: Qwen3.5 2B/4B/9B, Gemma 4 E2B/E4B/12B/26B, Qwen3.6 35B and Qwen3.8 27B. The chat offers "Add image support" the first time you attach an image.

Documents of any length. A long PDF no longer fails with "too long". The model's memory grows to fit it when your device allows; if it still doesn't fit, only the parts most relevant to your question are used, and the chat says so. Progress is shown while a long document is read.

Four more models. LFM2.5 1.2B for older phones, Gemma 4 E4B, Gemma 4 26B and Qwen3.8 27B for desktops with 32 GB of memory or a 16 GB graphics card.

Under the hood. 8-bit memory for the conversation, halving what a long chat costs; documents are refused at attach time only when they are of no use to the model (an image with no readable text, on a model that can't see).

Install

Download the file for your device from the assets below, or use the direct links in the README. Windows shows a SmartScreen warning because the app is not yet code-signed: More info → Run anyway. On a Mac, right-click the app → Open the first time.

opnlocal 0.1.0

opnlocal 0.1.0

First public release. opnlocal runs AI models on your own computer or phone, for people who
have never heard of VRAM or quantization. It checks what your device can handle, suggests up to
three models for what you want to do, downloads the one you pick, measures how fast it really
runs, and lets you chat. Your chats never leave the device.

Download

Platform	File	First launch
Windows 10/11 (x64)	opnlocal_0.1.0_x64-setup.exe	Not code-signed: SmartScreen warns → More info → Run anyway. No admin rights needed.
macOS 13+ (Apple Silicon)	opnlocal_0.1.0_aarch64.dmg	Not signed or notarized: right-click the app → Open.
Linux x64 (Ubuntu 22.04+)	opnlocal_0.1.0_amd64.deb	sudo apt install ./opnlocal_0.1.0_amd64.deb
Linux x64 (most distributions)	opnlocal_0.1.0_amd64.AppImage	chmod +x it, then run it. It bundles Ubuntu libraries (GTK, WebKitGTK and more); their licenses are inside, their source is in opnlocal_0.1.0_amd64.sources.tar below.
Android 9+ (arm64)	opnlocal_0.1.0_android-arm64.apk	Allow installing from this source. Needs a phone from about 2019 or newer.
iOS 17+	opnlocal_0.1.0_ios-unsigned.ipa	Sideloading only, e.g. with Sideloadly and your Apple ID. See RELEASING.md.

SHA-256 checksums are in SHA256SUMS.txt. The Linux, macOS and iOS files were built by this repository's
build workflow; the
Windows and Android files on the maintainer's PC (the APK is signed with the project's release key).

What's in it

• Hardware check: memory, graphics, processor, free space, and whether the AI engine can run.
• Four goals: everyday help, coding, writing, private documents.
• A signed catalog of 9 tested models: Qwen3.5 0.8B, 2B, 4B and 9B, LFM2.5 2.6B, Gemma 4 E2B,
  Gemma 4 12B, gpt-oss 20B, Qwen3.6 35B. Models with their own license terms ask you to accept
  them first.
• Downloads with progress, cancel, automatic retry and resume. Every file is checked against
  the catalog's sha256.
• A short speed test on your device, in plain words.
• Chat: saved conversations, Markdown and code, attach PDF, Word and text files, stop, try
  again, optional "Think harder".
• Settings → Advanced: graphics acceleration, context length, threads, import your own .gguf.
• Engine: llama.cpp in-process. Vulkan on Windows and Linux, Metal on macOS and iOS, CPU on
  Android.

Privacy

Network use is limited to model downloads you start (huggingface.co) and a signed model list
checked at most once a day (github.com; can be turned off). No accounts, analytics, telemetry
or crash reports. No local server or open port.

Tested, and not tested yet

• Windows: the installed build, full journey with real downloads, speed test and chat, on
  a Ryzen 7 7700X + Radeon RX 7800 XT. All 9 models verified. NVIDIA, Intel graphics and
  Windows 10 not tested.
• Android: full journey in the Android emulator. Not tested on a physical phone.
• Linux: CI machine: the .deb and the AppImage both detect hardware and generate text (CPU).
  Not tested on a desktop with a graphics card.
• macOS: CI virtual Mac: detects hardware, generates text with Metal. Not tested on a real
  Mac.
• iOS: iPhone simulator: detects hardware, generates text. Not tested on a physical
  iPhone.

Known limitations

• Nothing is signed by a recognized publisher yet (Windows, macOS, iOS).
• iOS apps signed with a free Apple ID stop opening after 7 days until re-signed, and may not
  get the extra memory that larger models need.
• Android runs on the processor only (no graphics acceleration yet).
• The app doesn't check for its own updates. Watch this repository's releases.
• Uninstalling on desktop leaves the data folder (models, conversations); delete it by hand.
  Settings → Advanced → Storage shows where it is.

License

Apache-2.0. Models have their own licenses, shown in the app before download.