Box v3.3.0
The biggest Box release ever: meet Box Assist — spoken camera assistance for blind and low-vision users — plus a rebuilt GGUF engine with real GPU acceleration, and on-device image generation.
🦯 Box Assist — One for the old folks..
A brand-new feature built for blind and low-vision users, living under the Core tab:
- Live mode — point the camera and Box calls out people, obstacles, and objects around you, with how close they are ("a person, very close, on your left").
- Reading mode — tap to have mail, labels, menus, and signs read aloud.
- Describe mode — tap and a vision model describes the scene in front of you in a couple of sentences, spoken aloud as it thinks — no long silent waits.
- Voice questions — double-tap, ask anything out loud ("what color is this shirt?"), and Box answers against what the camera sees. Double-tap again anytime to interrupt and ask something new.
- One download, everything included. The Box Assist download bundles the vision models, the Describe brain (Gemma-4-E2B), and the speech-recognition model (SenseVoice) — one progress bar with the true total size, and the whole feature works the moment it finishes.
- Hands-free camera smarts: continuous autofocus with a focus sweep before every read/describe (no tap-to-focus needed — or possible), automatic flashlight when it's too dark (Box tells you), and a blur check so Reading waits for a sharp frame instead of failing on a shaky one.
- Giant gesture controls (tap / double-tap / swipe / hold-to-repeat), physical volume-button controls (volume-down acts, volume-up switches modes — buttons you can find by feel), spoken + haptic feedback on everything, and TalkBack coexistence — Box Assist speaks with one voice, no double-talk.
- The screen never times out while Box Assist is open, and a long-press app shortcut on the launcher icon jumps straight into Box Assist.
- New Settings → Box Assist section: speech rate (or leave your TalkBack rate untouched), "announce hazards only," and haptic feedback toggles.
⚡ GGUF models: rebuilt engine, real GPU acceleration
The llama.cpp engine powering imported GGUF models got a ground-up overhaul:
- Full Vulkan GPU offload — flip the CPU/GPU chip in any GGUF chat and all layers run on your phone's GPU.
- Massively faster CPU mode — a flaw that routed CPU-mode prompt processing through the GPU is gone. On a Pixel 6a: 0.7 → 21 tokens/sec.
- Instant replies — model loading now does all the waiting: weights are read up front and reopened chats replay their history during the loading screen, so your first message starts generating immediately.
- Tokens/sec stat under every GGUF reply.
- New Settings → GGUF Models panel: context size, CPU threads, GPU layers, mmap, mlock, and a Q8 KV-cache toggle — defaults match previous behavior exactly.
- Sturdier model importing (files are copied and byte-verified before use) and automatic GPU→CPU retry if a model fails to load on GPU.
- llama.cpp updated to a current build with the newest Vulkan kernels and model support.
🎨
- FLUX.2 klein (4B) — photorealistic images in 4 steps. ~7.4 GB download.
- Z-Image Turbo — Tongyi's turbo image model, 9 steps.
🔍 Five new vision models — bundled in, work instantly
No download needed — these ship inside the app and run offline immediately:
- Identify now hosts four model families in one picker: MobileNet V2, MobileNet V3 Large (with a Pixel Tensor G5 NPU variant), PlantNet — identify 1,081 plant species from a photo — and DM-Count crowd counting.
- Erase (NEW tile) — paint over anything in a photo and MI-GAN inpainting removes it. Brush size, iterative "erase more," save to gallery.
- Upscale gains EDSR ×4 alongside the existing models.
Broader device support
- Minimum Android version lowered from 16 to Android 14
Both editions (Main + custom-rom-support) updated.