Skip to content

Releases

Bonsai 2 v3.0.0 - Offline Windows and Android

Choose a tag to compare

@alphastack1 alphastack1 released this 18 Oct 05:55

Downloadable Qwen3.5-based Bonsai 2 2B, 4B, and 9B models, with a shared interface for Windows and Android.

  • Bonsai-2-Windows.exe — single Windows x64 executable; CPU, Vulkan and NVIDIA CUDA engines included.
  • Bonsai-2-Android.apk — signed ARM64 APK for Android 9+. Same package and sideload signing certificate as the previous APK, version code 4.
  • Both include the smallest new model, Bonsai 2 2B, for offline use immediately after the initial local extraction.
  • Optional 4B, 9B, and matching image projectors download onto the device and then also run completely offline. No hosted inference, subscription, account or cloud API key.
  • Local chat/image history, model switching, stop generation, deeper thinking, resumable verified downloads, context and thread controls.

Checks

Short PC tests on the i5-12500H / RTX 4050 Laptop GPU (6 GB VRAM) / 16 GB RAM produced approximately 84 tokens/s with 4B and 54 tokens/s with 9B, thinking off. These are single short text samples, not phone measurements or a controlled comparison with the old app.

All six upstream model files (three language models plus three Q8 vision projectors) are backed up privately in alphastack1/Bonsai-2-GGUF-backup, with original revisions, SHA-256 values, attribution and Apache 2.0 licensing. The app downloads optional files from the pinned public upstream releases; private backup access is not required.

See the repository README and docs/validation.md for setup, memory expectations, build commands and detailed validation. SHA256SUMS.txt covers the two downloadable apps.

Bonsai Android v2.1.0 - Offline Multi-Model

Choose a tag to compare

@alphastack1 alphastack1 released this 18 Jul 18:43

What is included

  • A verified Bonsai 1.7B model embedded in the APK, so the app can work fully offline after first-launch extraction.
  • Optional verified downloads for Bonsai 4B and 8B text models, plus the 27B vision model.
  • Persistent model selection with safe switching and deletion controls.
  • Advanced controls for CPU threads, context size, system prompt, temperature, Top-P, Top-K, and response length.
  • A redesigned Android launcher icon and in-app bonsai mark.

Existing v2.0 installations keep their app data and previously downloaded 27B model when updated normally. Do not uninstall the old version first if you want those files preserved.

Install and storage

This phone-only build requires Android 9 or newer on an ARM64 device. Download the APK, allow installation from your browser or file manager when Android asks, and install it as an update.

The APK is about 260 MB. First launch extracts the embedded 1.7B model (about 248 MB), so allow roughly 500 MB total for the app and offline starter model. Optional models need additional free space plus temporary download headroom.

This sideload release is signed with the same project debug signing key as v2.0.0 and has a higher version code, allowing an in-place update from that release.

Verification

  • APK SHA-256: 9E2C7C7E63205F5B85807DC9DAA56F66ECBA13A094091639F73745B6B699BE54
  • Release build, Android lint, APK alignment, v2 signature, embedded model contents, and phone-sized UI/model/settings flows were checked locally.
  • The v2.1.0 APK itself still needs the final physical-phone smoke test. The earlier v2.0.0 27B build was confirmed running on a Pixel 9 Pro XL, at roughly 0.5 tokens/second.

Because this repository is private, the phone must be signed into a GitHub account with access before downloading the asset.

Bonsai 27B Android v2.0.0

Choose a tag to compare

@alphastack1 alphastack1 released this 18 Jul 17:55

Bonsai 27B Android v2.0.0

Phone-only, arm64 Android sideload release of the local multimodal Bonsai 27B app.

What is included

  • A single 12 MB signed APK containing the UI and the pinned PrismML llama.cpp runtime.
  • First-run resumable downloads of the official 3.80 GB language GGUF and 629 MB vision projector.
  • Exact size and SHA-256 verification before downloaded weights are accepted.
  • Local text and image chat; selected photos are resized on-device and are not uploaded.

Verification performed

  • Clean release assembly succeeded for arm64-v8a.
  • Android lint completed with zero errors.
  • APK alignment, v2 signature, package metadata, native ABI, and absence of embedded GGUF weights were checked.
  • Phone-sized mocked UI flows covered setup/download/retry, model load/switch/delete, text chat, image chat, stop, clear, copy, sidebar, settings, and error states.

An actual Android handset was not connected during this audit, so full on-device model loading, memory use, and inference speed remain the purpose of this test release.

Install

Download Bonsai-27B-Android-v2.0.0.apk on an Android 9+ arm64 phone, allow installation from that source, and open the APK. In the app, open Models and download the 4.43 GB verified model bundle. Keep about 5 GB free for the weights plus download headroom; a 12 GB-class phone is recommended.

SHA-256: 801A13AD08F21263196075A1AF6DDF1D2EA1714BE9CBB1EA294EA99A9198EAC1

This is a sideload test build signed with the standard Android debug certificate, not a Play Store release.