Box v3.3.1
The most advanced local AI on Earth. State-of-the-art models running entirely on your phone — no cloud, no account, no connection required.
Meet the Live Translator — two people, two languages, one phone — plus pausable downloads, four new models, new options, and a wave of fixes from the 3.3.0 issue's. Everything runs entirely on your phone, offline.
Live Translator (NEW)
A new tile under the Sound tab: hand the phone between two people speaking different languages and Box interprets both ways.
- Tap your button, speak, done. Recording stops by itself when you pause; the translation appears on screen and is spoken aloud in the other person's language.
- Auto conversation mode — flip one switch and Box keeps the conversation going: after each translation is spoken, it automatically listens for the other person. No taps between turns, and the screen stays awake.
- 24 languages, with a swap button and your picks remembered.
- Tap any translation to hear it again, or copy it with one tap. A Speak toggle switches to text-only for quiet places.
- No new download — it runs on the Gemma audio model you already use for chat (E2B/E4B). The model hears the speech directly; there's no separate speech-to-text step.
Downloads you can pause now
- Pause and resume model downloads — a new pause button on every downloading card. Paused downloads survive closing the app and pick up exactly where they stopped, even for multi-gigabyte models.
- The download notification now shows percentage, speed and time remaining, with a Pause action right on it.
- New Settings → Downloads & Updates → "Download models on Wi-Fi only" (off by default) — when on, big downloads wait for Wi-Fi instead of using mobile data.
🤖 Four new models
- Granite 4.0 350M — IBM's tiny, fast instruct model (468 MB). Great on any device.
- MiniCPM5-1B — 1B-class open-source SOTA, strong in English and Chinese, in two builds: int8 (GPU-optimised) and a smaller int4.
- Gemma 4 26B (A4B) — experimental: Google's mixture-of-experts Gemma 4, 26B parameters with 4B active per token. Text-only, ~15.8 GB, requires 16 GB+ RAM — this is Google's web build running ahead of official mobile support, so expect rough edges.
⚙️ New options in Settings
- Chat → "Show tokens per second" — see a decode-speed stat under every chat reply, for all models (GGUF replies already had it).
- Chat → "Remember reopened chats" — choose between full conversation memory when reopening a chat (default) or instant first sends.
- Box Assist → "Live callout amount" — tune how chatty Live mode is: More / Balanced / Fewer-but-surer.
- Downloads & Updates → "Check for updates automatically" — control the update banner; the manual check works regardless. The automatic new-release banner also got fixed — it was checking the wrong place and could never appear.
🛠 Fixes — thank you for the reports!
- GGUF models no longer crash on GPU on some Snapdragon devices (#109, #119). The Adreno driver advertises a Vulkan feature whose entry point it never provides; loading any GGUF with the GPU chip enabled crashed instantly. Box now checks the entry point actually exists before using it.
- Rotating or unfolding your phone no longer unloads the model (#111). The screen used to go white and your chat model had to reload — now the app just reflows.
- Custom-ROM edition: TPU/GPU chat works again on de-Googled devices (#112, GrapheneOS). Chat now loads NPU-built models text-only on those devices (matching how the benchmark always worked) instead of failing with "Input tensor not found". Image/audio input with the Tensor G5 build stays unavailable there — use the standard GPU build of Gemma E2B for those.
- Z-Image generation failures fixed (#118 and friends): any generation graph that can't compile on your GPU now runs on CPU automatically instead of failing the whole image. Applies to FLUX.2 klein too.
- Failed-download notifications now say the download failed (they previously showed the success message).
- Accurate model info: the Gemma 4 Tensor G5 card's "32K context" and the 4000-token default are both real — 32K is the model's maximum, 4000 is Box's conservative default session budget (#116).
- various other misc bug's + another round of security audits + hardening.
⚠️ Known issue
- Google's official gemma-4 QAT Q4_0
.gguf(from the QAT GGUF repo) ships with a broken vocabulary upstream — three duplicate tokens — and can misbehave badly when imported, up to freezing the phone with GPU enabled. Until it's fixed upstream, use Box's built-in Gemma 4 models (.litertlm) instead of importing that file. A load-time repair is planned
Some GGUF models wont work on GPU acceleration.
- Box now ships with 70 + models available.
Huge thank you to everyone for suggestions,bug reports ! and a Huge massive thank you to LiteRt Community and everyone making good quality model's ! If you are a author of a model featured in Box and want to be on the readme let me know <3 Box would not be possible without the great community behind Box, very much appreciated ! <3
Much much more coming soon.
Both editions (Main + custom-rom-support) updated.