TopGit
GitHub Repo Review

Box: Offline AI Android App for On-Device Chat

jegly/Box
BTopGit review image for jegly/Box
Review by Topgit.dev for jegly/Box, with GitHub repository stats and README context.
Quick verdict

Box is an offline AI Android app running chat, image generation, vision, voice, and music generation entirely on your phone, no server round-trip. It forked Google's AI Edge Gallery; its README says upstream merging has stopped, with Box now carrying more than 50 features the original lacks. Reach for it over five separate apps, especially on a de-Googled phone; skip it for a single fast chatbot, since setup asks more than a Play Store app.

Stars
★ 833
Forks
⑂ 50
Language
Kotlin
License
See repository
Topic
AI Tools
Updated
Sep 2026
Homepage
GitHub

What is Box?

Box is an offline AI Android app that packages on-device chat, image generation, vision, voice, and music generation into one install, working without internet once its models are on the phone. The README calls it a fork of Google's AI Edge Gallery that has since diverged substantially, adding face recognition and a spoken camera mode for blind and low-vision users, via GPU, TPU, or NPU acceleration depending on your phone's chip.

Key On-Device AI Capabilities

  • GGUF chat runs on a rebuilt llama.cpp engine with full Vulkan GPU offload — the README cites a jump from 0.7 to 21 tokens per second on a Pixel 6a after the v3.3.0 rewrite, plus a Settings → GGUF Models panel for context size, CPU threads, GPU layers, mmap, mlock, and Q8 KV cache.
  • Three on-device image-generation engines: Bonsai Image 4B — a ternary-weight take on FLUX.2 klein, about 4.3 GB to download, CPU-only, producing 512×512 images in four steps — plus FLUX.2 klein (about 7.4 GB) and Z-Image Turbo (about 10.6 GB, nine steps), with EDSR ×4 upscaling to push results to 2048×2048.
  • Box Assist gives blind and low-vision users a spoken camera mode with Live, Reading, and Describe modes, plus voice questions answered against whatever the camera is pointed at.
  • The Sound tab covers speech-to-speech voice chat, a Live Translator running on the Gemma audio models across 24 languages, multilingual text-to-speech, and three tiers of music and sound generation (SoundGen, SoundGen HD, SoundGen HD Long, up to about three minutes).
  • Bundled vision tools include object and scene recognition across several model families (PlantNet alone covers 1,081 plant species), on-device face recognition, and an Erase tool built on MI-GAN inpainting.
  • Acceleration is matched to your exact chip: GPU everywhere, Google's Tensor G5 TPU on Pixel 10, and NPU dispatch libraries tuned for specific Snapdragon and MediaTek chipsets, with Box picking the right build automatically.
  • Security features are named specifically rather than implied: biometric app lock, encrypted chat history, and tap-jacking protection, alongside a separate custom-rom-support build for GrapheneOS, LineageOS, and CalyxOS.
  • GGUF model import lets you load your own local model instead of relying only on Box's bundled catalog.
How this repository's GitHub stars have grown over time. Source: star-history.com.View the star history

Practical Applications of Box

  • Running local LLM chat against GGUF or Gemma models with no signal — on a flight, underground, or anywhere offline.
  • Using Box Assist as a spoken navigation aid: Live mode narrates obstacles and people, Reading mode reads mail or menus aloud, for a blind or low-vision user without a cloud vision API involved.
  • Generating a quick sound effect or short music clip for a video edit through SoundGen, without uploading audio to a paid cloud generator.
  • Running a full on-device AI suite on a de-Googled phone via the custom-rom-support build, where most Play-Services-dependent AI apps don't work at all.
  • Feeding a document into Box for offline Q&A, or asking Ask Image to read text out of a photo with the InternVL vision models.

Installing Box on Your Android Device

Box's README documents an Obtainium-based install instead of a Play Store listing: add the repo URL (https://github.com/jegly/Box) inside Obtainium's add-app flow, and it fetches whichever release track you choose. Two tracks exist — Main, built for stock Android on phones like Pixel or Samsung, and Custom ROM, aimed at GrapheneOS, LineageOS, or CalyxOS devices that don't run Google services — and Obtainium needs the corresponding filter value (Main or custom-rom-support) set so it grabs the right APK. From there, new releases show up in Obtainium automatically, or you can check by hand from Settings inside Box. The README notes the in-app version number has tracked the actual Box release number since v2.0.0, correcting an earlier mismatch with the build number Box inherited from Google AI Edge Gallery. The changelog also states the minimum supported Android version is Android 14.

Strengths

  • Feature breadth in one APK: chat, image generation, vision, voice, and music generation, without juggling five separate apps.
  • Concrete engineering behind the acceleration claims — the README's own before/after numbers (0.7 to 21 tokens per second on a Pixel 6a after the GGUF engine rewrite) are the kind of specific detail generic 'fast AI' marketing usually skips.
  • A dedicated custom-rom-support build and setup path for GrapheneOS, LineageOS, and CalyxOS, a case most mainstream Android AI apps ignore entirely.
  • Security features named specifically rather than claimed generically: biometric lock, encrypted chat history, and tap-jacking protection.
  • An active, detailed changelog — version-by-version entries down to specific issue numbers (#67, #81, #82, #83, #88) suggest bugs get tracked and closed rather than silently dropped.

Current Limitations and Considerations

  • The repo's listed license field is blank ('?'), so don't assume redistribution or contribution terms — check the actual LICENSE file on GitHub before relying on it.
  • Diffusion image generation isn't fast on typical hardware: Bonsai Image is CPU-only because its diffusion graph is too large for the GPU delegate, and the README puts render time at roughly two minutes per image with memory use near 8 GB.
  • NPU acceleration is chipset-specific, not a blanket 'Android' feature — the changelog logs a string of NPU crash fixes tied to particular Snapdragon and MediaTek SoCs (e.g., SM8750, SM8750-AB), so results depend on your exact phone model.
  • In-app language coverage is partial: the Settings → Language picker added French and Portuguese alongside English, but only some screens are translated so far and the rest stays in English.
  • Box no longer actively merges changes from upstream Google AI Edge Gallery, per its README, so any fixes or features Google ships there arrive in Box only if its own maintainers port them over.
  • It's a young project by GitHub metrics with no Play Store listing documented — installation runs through Obtainium and GitHub releases only, a higher bar than a typical Android app install.

Alternatives to Box

Google AI Edge Gallery — the official Google project Box forked from; the standard on-device demo app, without Box's custom-ROM build or the more than 50 extra features the README says Box has added since.PocketPal AI — a lighter, chat-only local LLM app for Android if you don't need image generation, voice, or vision tools.ChatterUI — another GGUF-focused Android chat client for people who want a simpler llama.cpp front end than Box's full suite.Ollama — runs local models on desktop or server rather than a phone; the closest analog if you want the same no-cloud approach off Android.LocalAI — a self-hosted, OpenAI-API-compatible server for local models, better suited to running behind your own backend than to a phone app.

Frequently Asked Questions

Does Box support custom Android ROMs?

Box ships a dedicated custom-rom-support APK built for GrapheneOS, LineageOS, and CalyxOS, separate from the Main version for stock Android — grab it from the latest GitHub release instead of the standard build.

Can Box run AI models completely offline?

Box runs its chat, image generation, vision, voice, and music generation features entirely on-device, according to the README, with no server round-trip once a model is downloaded to your phone.

What kind of hardware acceleration does Box use?

Box routes chat and image generation across GPU, Google's Tensor G5 TPU on Pixel 10, and NPU dispatch libraries tuned for specific Snapdragon and MediaTek chipsets, picking the build that matches your phone's chip automatically.

Can I import my own GGUF models into Box?

Box supports GGUF model import, per its README, so you can load your own llama.cpp-compatible model instead of only using the ones bundled in Box's download catalog.

What languages does the Box app support?

Box's Settings → Language picker covers System, English, French, and Portuguese, including Português (Brasil), though the changelog notes translation coverage is partial and some screens stay in English.

Is Box affiliated with Google?

Box began as a fork of Google's AI Edge Gallery, but its README states it is not affiliated with or endorsed by Google LLC, and Google branding has been replaced throughout the app.

The problem it solves

On-device AI apps for Android tend to be single-purpose — one app for local chat, another for image generation — and most assume stock Android with Google Play services, leaving GrapheneOS, LineageOS, and CalyxOS users with fewer working options. Box's README frames its own origin around that gap: it forked Google's AI Edge Gallery and then built out a separate custom-rom-support release track for de-Googled phones, while folding chat, image generation, vision, voice, and music generation into one app instead of five separate ones.

Who should try it — and who should skip

Try Box if you're on Android and want one app for offline chat, image generation, vision, voice, and music generation without a network call — especially if you run GrapheneOS, LineageOS, or CalyxOS and most AI apps simply don't install for you. Skip it if you just want a single fast chatbot: a narrower app will be quicker to set up, since Box's Obtainium-based install and per-chip acceleration tuning ask more of you upfront than a Play Store download.

Related repositories

Source & attribution

Facts and figures sourced from the jegly/Box GitHub repository, its README, and its version changelog.

GitHub data · last synced Aug 12, 2026Reviewed by Henry
Back to TopGit

Still deciding about Box?

One click hands the question to an AI along with this page — see what it says about Box.

GitHub