Onionmind desktop (rolling)

Rolling build of main, republished on every push. The installers are attached here because a raw link renders in the browser instead of downloading; release assets are served as attachments. Installed copies check this release over Tor; the manifest pins the revision, size, and SHA-256 of the zip. SHA256SUMS covers every asset and carries a GitHub build-provenance attestation (gh attestation verify SHA256SUMS --repo Codemaster64/onionmind).

Onionmind v1.4

The Android app gets the workbench v2 features that fit a phone, and the Nocturne look.

New (Settings → Features). Every feature starts off except "Ask before searching the web":

• Ask before searching: asks before a Tor search; your answer counts for that turn only.
• Reply details: where a reply came from (on this phone, which model, Tor search results and domains), whether reasoning was hidden, and how long it took.
• Live stats: the last reply's speed and free memory, plus a context-window meter that uses the model's own counts.
• Models: fit labels for this phone's memory, and switching loads the new model straight away.
• Network: a boundary map and a network log (route and host only, never what you asked) that you can share as JSON or CSV.
• Chat tools: edit and resend, ask again, presets, and starter prompts.
• Recovery: error cards with fixes for a model that won't start, low memory, and a failed Tor search.
• Privacy: auto-delete the conversation after 1, 7 or 30 days, and a private screen that blocks screenshots and the app-switcher preview.
• Look: Nocturne colors and an optional light theme.

Models: SPARK (LFM2.5 2.6B), EMBER (Qwen3.5 4B) and BLAZE (Qwen3.5 9B), now described the same way as on the desktop.

Nothing here changes the privacy boundary. Search is still Tor-only, per turn and fail-closed, and there is no new network path. Details: docs/android-v2.md.

Install notes

• Uninstall the old version first. Like every earlier release, this APK is debug-signed with a different key, so it won't install over v1.3. Your conversation and downloaded models are removed with the old app, so the models have to be downloaded again.
• Not yet tested on a phone. The Kotlin tests pass and the UI was tested in a phone-sized browser, but no device or emulator has run this build.
• Build: arm64, Android 8.0+. Built from 628f7ff with llama.cpp f2cc728.
• SHA-256: ee3cc05fdf85d0353cea04b1143a0f27df0dd042ac1e2cf4f9cbaa472fc50286

Onionmind v1.3

Transparent launcher icon background; preserves alpha in legacy icon resources.

Onionmind v1.2

Adds the latest mobile model picker with LFM2.5, Qwen3.5 4B/9B, and live download progress.

Onionmind v1.1

Fix Android model download startup, resumable transfers, and model-size completion handling.

Onionmind 1.0 - one-click everything

One download per platform, and every install gives you the onionmind terminal command.

File	What it is
Onionmind-Setup.cmd	Windows, one double-click. The entire installer in one self-extracting file.
install-onionmind.sh	Linux - download, then bash it.
Onionmind-0.1.apk	Android - tap to install; model downloads on first launch.
install-onionmind-android.sh	Android the tinkerable way (Termux, gives the onionmind command).

Matchstick ISOs are not here: GitHub caps release files at 2GB and the images are 5.5-16GB. Build them from source - one docker command, see usb/README.md.

Matchstick Pocket - prebuilt (spark edition)

The pocket Matchstick stick, pre-built - no Docker, no source checkout needed.

4 parts (GitHub caps files at 2GB): the matchstick.cmd / matchstick.sh script downloads and rejoins them automatically, or manually: download all 4 parts + SHA256SUMS, cat onionmind-matchstick-4b-amd64.iso.part?? > onionmind-matchstick-4b-amd64.iso (Windows: copy /b), verify against SHA256SUMS, burn with matchstick.cmd / matchstick.sh / Rufus.

Boots on any x86-64 PC, Secure Boot OFF. NVIDIA GTX 900-RTX 40 or modern AMD for GPU speed; anything else runs CPU. First boot = the chat. 32GB NOT required: 16GB stick.