Box v3.0.0

A major UI overhaul — the whole app got a fresh, expressive, fully theme-aware coat of
paint, plus a big stability fix for Snapdragon and MediaTek NPU devices. The jump to 3.0.0
reflects how much of the interface changed, not a new engine: everything you loved still runs the
same models the same way, it just looks and moves a lot better now.

Themes & personalization

• 11 new themes inspired by the GNOME Ptyxis terminal palettes — Fairy Floss, Nord, Bim,
  Borland, C64, Cobalt Neon, Grass, Homebrew Ocean, Mono Amber, Mono Red, and Synthwave —
  selectable from a new palette dropdown in Settings. (Joins the existing System, Light,
  Catppuccin, and Dracula themes.)
• Choose your app font. 13 bundled font families — Nunito (default) plus Cormorant Garamond,
  DotGothic16, IBM Plex Mono, IBM Plex Serif, Instrument Serif, Egyptian Hieroglyphs, Playfair
  Display, Press Start 2P, Quicksand, Space Grotesk, Turret Road, and Viaoda Libre — each
  previewed in its own typeface.
• Separate chat font. Pick a different font just for chat messages, or leave it following the
  app font.
• Text-size slider. Scale text across the whole app (and chat) from 0.8× to 1.4×, on top of
  your system font size.
• Themed app icon. With "Themed icons" enabled, the Box launcher icon now tints to your
  system Material You colors.

A more expressive interface

• M3 Expressive motion throughout. Spring-physics animations bring the app to life — home
  cards bounce in and respond to taps, chat messages rise and fade in, and screen transitions use
  Material 3 slide-and-fade.
• Settings, reorganized. The long settings list is now grouped into collapsible categories —
  Appearance, Privacy & Security, Network & Tools, Chat & Voice, and About — that expand and
  collapse smoothly.
• Cleaner home screen. Task cards are now clean and text-focused (the mismatched per-task
  icons were removed), the side-drawer cards were simplified, and the menu button is now a tidy
  box/cube glyph that suits the app's name.
• Theme-aware task screens. Open any task (Chat, Diffusion, Voice, …) and the background now
  follows your selected theme with the same subtle accent tint as the home screen — no more flat
  black behind a colorful theme. The oversized icon above each task title was removed for a
  cleaner header.
• New download progress style. Model downloads now show the Material 3 Expressive wavy
  progress indicator.
• Better contrast on themed tabs. The selected tab highlight now follows your theme's accent
  color (instead of turning hot pink on some palettes), with readable text on top.

Fixes & stability

• NPU crash fixed on Snapdragon & MediaTek (#82, #83). NPU models could hard-crash on load on
  non-Pixel devices (e.g. Galaxy S26 Ultra, Xiaomi 14T Pro) because the wrong hardware dispatch
  library was being loaded. Box now selects the correct Qualcomm / MediaTek runtime per device.
  The Pixel 10 / Tensor G5 path is unchanged and was re-verified.
• No more Settings flash. Changing the text-size slider no longer flashes the home screen
  behind Settings; font scaling is now applied cleanly via typography.
• Smoother Settings dialog. Fixed the jump/jank when opening Settings or expanding categories
  — the panel is now a fixed size and animates its contents internally.

Build variants

• Main (stock Android): includes the Gemini Nano Hub & Vision features (AICore modes need a
  supported device).

• custom-rom-support (GrapheneOS / LineageOS / CalyxOS): no Google-services-dependent features
  required.

Both variants include the full UI overhaul, all new themes and fonts, and the Snapdragon /
MediaTek NPU fix.

Notes for testers

• The Snapdragon (sm8850) and MediaTek (mt6989) NPU fix is confirmed inert on Tensor G5 (no
  regression) but still needs on-device confirmation on those Snapdragon/MediaTek phones.
• Per-SoC NPU model builds added in 2.0.2 remain untested on their target hardware.

Box v2.0.2

New models

• Gemma 3 270M — a brand-new ultra-lightweight tier (~460–555 MB). Fast, low-RAM, with dedicated NPU builds for Snapdragon and MediaTek.
• Gemma 3 1B-IT — now with broad on-device NPU coverage:
  • Snapdragon (8 Gen 1 → 8 Elite): SM8550, SM8650, SM8750, SM8750-AB, SM8850
  • MediaTek Dimensity: MT6989, MT6991, MT6993
  • plus a universal GPU/CPU build for everything else
• Gemma 3n E2B & E4B — multimodal (text, image, audio), up to 32K context. Runs on GPU/CPU on all devices; NPU acceleration available on MediaTek (MT6993).

Device support

• Samsung Galaxy S26 Ultra (Snapdragon 8 Elite Gen 5 / SM8850) — NPU models now appear and run correctly. Adds SM8850 keys across the new Gemma 3 1B and 270M entries.

• All new models are subject to Google's Gemma Terms of Use and Prohibited Use Policy.

Box v2.0.1

Bug fixes for Snapdragon NPU users

Fixes

Snapdragon 8 Elite NPU crash (all SM8750 / SM8750-AB devices)

The audio sub-graph was incorrectly routed to the NPU when loading the Snapdragon NPU model, causing an instant hard crash (SIGABRT) with no error popup on every affected device.
Audio always uses CPU regardless of the primary backend —
this matches upstream behaviour and is how the NPU .litertlm files are compiled.
Affects Red Magic NX799J, iQOO 13, and all other SM8750 and SM8750-AB devices.

Samsung Galaxy S25 / S26 Ultra — NPU models not listed

Samsung's "Snapdragon 8 Elite for Galaxy" variant reports SM8750-AB as its SoC identifier rather than SM8750.
Box 2.0.0 only matched sm8750 in the model list, so the dedicated Snapdragon NPU model was invisible on all S25 and S26 Ultra
devices. The SM8750-AB key is now included and maps to the same binary.

New model

Gemma 4 E2B — Qualcomm QCS8275 (Dragonwing IQ8)
Added an NPU model entry for the QCS8275 SoC.

Box for Linux v0.3.0

A polish + stability update for Box for Linux — the native desktop AI assistant that runs models fully locally on your machine.

What's new

• RAG / knowledge base — attach files to a chat and the assistant retrieves relevant context automatically. Supports text, PDF, and JSON. Index documents into shared notebooks that can be attached across multiple chats. Vector search powered by TurboVec (4-bit quantised ANN index).
• Thinking indicator — a colourful animated thinking label appears while the model is generating, so it's always clear when Box is working.
• Hover actions on messages — copy, speak, or save a message as a memory by hovering over any bubble. Buttons appear on hover and don't shift the layout.
• Copy conversation — copy the full chat as plain text with one click (button in the composer bar).
• Notebook navigation — back button in the notebook view; switching to the Chats tab also returns to the chat pane.

Also fixed

• Context overflow crashes (Input token ids are too long) when sending messages in chats with long RAG history — history is now trimmed automatically to fit the context window.
• RAG context overflow on JSON-heavy datasets — retrieval budget now correctly accounts for token-dense content.
• GPU model error dialog now warns explicitly not to switch to GPU on integrated graphics.
• Context window preference subtitle corrected (was "Default 32k", now "Default 4096").

Install

sudo apt install ./box_0.3.0_amd64.deb

Dependencies are pulled automatically. Launch Box from your app menu or run box. Upgrading from v0.2.0 installs over the top — your chats, models, notebooks, and settings are kept.

Requirements

• Ubuntu (amd64) with a GTK4 / libadwaita desktop session
• Python 3.14
• ~3–4 GB free storage for a model
• CPU-only works fine; GPU acceleration is faster but optional

https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression/

Checksum

Box

Box's first major-version release. The headline is hardware NPU/TPU acceleration coming to
life — the Google Tensor G5 (Pixel 10) and MediaTek neural engines now run models on-device,
joined by a batch of new models, an on-device Face Recognition feature, and a brighter,
fully theme-aware interface.

⚡ Hardware acceleration

• Google Tensor G5 (Pixel 10) acceleration (new). Gemma now runs on the Pixel 10's Tensor
  G5 neural engine, not just the GPU. Supported models route to the TPU automatically on Pixel 10
  and expose a dedicated TPU option in the accelerator picker.
• MediaTek NPU support (new). Box now ships the runtime and the first models that run on
  MediaTek Dimensity neural engines — the first MediaTek NPU models in Box.
• Snapdragon & Tensor NPU packaging fix. Native NPU/DSP libraries are now correctly unpacked
  on install, so NPU models initialize reliably instead of failing to find their runtime.

🤖 New models

• Gemma 4 E2B (Tensor G5) — NPU-optimized build for the Pixel 10. Multimodal (text + image +
  audio), 32K context, thinking.
• Gemma 4 12B — large Gemma 4 model (GPU), text, 32K context. For high-memory devices.
• Gemma 3 1B-IT (Tensor G5) — compact Pixel 10 NPU model.
• Gemma 3n E2B (MediaTek) — multimodal (image + audio) NPU model for MediaTek devices.
• Qwen3 0.6B (MediaTek) — compact MediaTek NPU model.

🧑 Face Recognition (new, on-device & private)

• A new Face Recognition feature (in the Image section): detect, enroll and name people, then
  recognize them in photos or live from the camera — fully offline.
• Enroll multiple samples per person for better accuracy; faces are aligned before matching for
  reliable results.
• Capture-to-add from the live camera, an on-screen face mesh overlay, and a settings
  panel (match strictness, front-camera default, show match %, clear all).
• All face data is encrypted on-device and never leaves the phone — opt-in and user-enrolled
  only, with no external or web-scraped identity database.

🎨 Interface & themes

• New Light theme. A crisp, truly-bright light mode that doesn't tint toward your wallpaper —
  for people who found the dynamic light theme not light enough.
• Theme-aware home screen. The home background now follows your selected theme (System /
  Light / Catppuccin / Dracula) instead of always being black, with a subtle accent tint so each
  theme is visibly distinct.
• Fixed the theme picker squeezing the "Catppuccin" label.

🤝 Gemini Nano (stock & custom-ROM)

• Gemini Nano Hub now on the custom-ROM build too — Summarize / Proofread / Rewrite /
  Describe / Chat / Speech, previously stock-only. (Degoogled devices without AICore degrade
  gracefully; ML-Kit vision features still work.)
• Fixed a crash when attaching a document in Summarize/Proofread/Rewrite (the file picker
  could be hijacked into the photo picker on Android 14+). Documents now open through the proper
  document picker, with a clean fallback if none is available.
• Fixed memory/service leaks when moving between Nano features.

💬 Chat improvements

• Copy button on code blocks. Fenced code blocks in chat now render with a language label and
  a one-tap Copy code button.
• Speech-to-text with SenseVoice in Chat. The chat mic now works with a loaded SenseVoice
  model (priority: Whisper → SenseVoice → system) — no longer dead-ends asking for Whisper.
• Speculative decoding is available for Gemma 4 in chat (off by default).
• Gemma Terms of Use. A one-time consent dialog (linking the Gemma Terms of Use and Google's
  Prohibited Use Policy) now appears before downloading any Gemma model.

🛠️ Fixes & under the hood

• Agent mode now works with text-only models (#69). Agent mode no longer force-loads vision
  on models that don't support it, which previously blocked text-only imported models entirely.
• Correct installed version (#67). Fixes Obtainium / "App version" showing the wrong version
  (and "update available" loops) by aligning the internal and public version numbers — the reason
  for the 2.0.0 jump.
• Large GPU models fall back to CPU instead of dead-ending when they exceed the GPU's limits.
• Gemma 4 12B initialization fix — loads correctly on GPU instead of failing with a
  misleading backend/audio-adapter error.

Build variants

• Main (stock Android): includes the Gemini Nano Hub & Vision features (AICore modes need a
  supported device; ML-Kit vision runs broadly).
• custom-rom-support (GrapheneOS / LineageOS / CalyxOS): now also includes the Gemini Nano
  Hub; no Google-services-dependent features required.

✨ Highlights

A big release: two new voice models, on-device image upscaling, a Gemini Nano vision upgrade,
a much-improved Markdown/LaTeX renderer, new language models, and a pile of fixes.

🎙️ Voice & Audio

• SenseVoice — fast multilingual speech-to-text (new). A new card in the Voice tab. Transcribes
  Chinese, English, Japanese, Korean and Cantonese, runs fully offline, and is roughly 5× faster
  than Whisper on CPU. Includes a live "listening" preview while you talk, a multi-message transcript
  log (copy/delete/clear), a language picker, punctuation/number formatting, and optional emotion &
  audio-event tags.
• Supertonic — fast multilingual text-to-speech (new). A new card in the Voice tab. Lightweight
  (~66M params) on-device speech synthesis in English, Korean, Spanish, Portuguese and French, with
  multiple built-in voices and adjustable speed. Fully offline — text never leaves your device.

🖼️ Image

• AI Image Upscaling / Super-Resolution (new). A new "Upscale" feature in the image tab. Pick any
  photo, enhance and enlarge it 4× on-device, and save the result to your gallery — fully offline.
  Choose between XLSR (fastest, tiny), Real-ESRGAN General (balanced) and Real-ESRGAN x4plus
  (highest quality). All three models are bundled in the app — no download required. Photos are
  auto-rotated correctly before upscaling.
• Gemini Nano Vision: visual overlays (stock build). Pose detection now draws a skeleton overlay
  and Face Mesh draws a 468-point mesh directly on the camera preview and still images (previously
  text-only). Plus: copy buttons on every vision result, an adjustable live refresh rate
  (Fast/Balanced/Slow/Power-saver) with a Freeze/Resume toggle, front/rear camera switching on
  all modes, and image upload from your gallery.

🤖 Models

• New language models: TinyLlama 1.1B, Phi-4-mini, TinySwallow 1.5B, VibeThinker 1.5B, and Qwen3 8B.
• Models browser organized by type. The model list is now grouped into Language models /
  Speech-to-Text / Text-to-Speech / Image generation / Other instead of one flat list.
• Clearer model guidance. Gemma 4 E2B is labelled "Recommended" and E4B "Best overall for
  flagship devices," with cleaned-up, less cluttered model descriptions.

✍️ Markdown & LaTeX rendering

• Major rendering fixes (#42). Headers, bullet/numbered lists and bold text now render correctly
  even when mixed with inline math on the same line, bold that spans a math expression no longer
  shows literal **, and wide display equations scroll instead of being clipped.

🧹 Interface

• Removed promotional banners/links from the MCP and Agent screens for a cleaner first-run experience
  (sample-prompt chips kept).
• Saving an upscaled image now shows a "Saving…" indicator and writes safely — leaving the screen
  mid-save no longer produces a cropped/partial image in your gallery.

🛠️ Fixes & under the hood

• Fixed a crash on some Snapdragon NPU devices when using vision/audio models (#59).
• Leftover model files are cleaned up after app updates (#61).
• Fixed a speech-recognition hang on GrapheneOS (custom-ROM build, #65).
• Fixed a crash opening the settings dialog on small-context-window models.
• Fixed the notification tap target (deep link).
• Updated to Android SDK 37.

Build variants

• Main (stock Android): includes the Gemini Nano Hub & Vision features (needs AICore / Pixel 9+ for
  AICore modes; ML-Kit vision runs on any device).
• custom-rom-support (GrapheneOS / LineageOS / CalyxOS): SenseVoice, Supertonic and Upscaling are all
  included; no Google-services-dependent features.

Box v1.0.11 — Release Notes

🌟 What's New

MCP (Model Context Protocol) experimental support in Agent

Connect your on-device model to external MCP tool servers (e.g., GitHub repo docs via gitmcp.io, custom internal servers). Per-call permission prompts (Always allow / Allow once / Don't allow), tool-call history in chat, and
configurable per-server tools.

• Master kill switch toggle in Settings (default OFF for privacy)
• First-enable disclaimer dialog
• Hard Offline Mode takes precedence: disables MCP when on
• Optional Request Header auth for private servers
• Long tool responses auto-truncated to fit the model's context window

Renamed: "Agent Skills" → "Agent"

The home-screen card is now simply Agent, with a new subtitle: "Complete agentic tasks with skills and MCP tools."

Chat layout restored to single-column (ChatGPT-style)

Both your messages and the model's responses are left-aligned with sender labels above each bubble. No more right-aligned user messages.

Refreshed model files

• Gemma 4 E2B / E2B (Snapdragon 8 Elite) — updated to latest litert-community release
• Gemma 4 E4B — same refresh
• Pulls in upstream model fixes and improvements

⚡ Hardware & Performance / UI

• Small UI tweaks / new themes.
• Pixel 8 / 9 / 10 now show "TPU" label for Tensor accelerators (was Pixel 10 only)
• Broader NPU init fallback — when an NPU backend fails to start for any reason (vendor mismatch, HTP allocation, format incompatibility), Box silently falls back to GPU. Addresses the iQOO 13 / Vivo OriginOS crash report. (Main APK
  only — custom-rom keeps the narrower previous check.)
• Streaming markdown rendering smoother — new BufferedFadingMarkdownText reduces re-renders during model output
• Faster chat scroll — ChatPanel switched to snapshotFlow + derivedStateOf, fewer recompositions during streaming
• MTP / speculative decoding stability fixes from upstream cherry-picks

🔒 Security & Privacy

• APK tamper detection — Box now verifies its signing certificate at launch and refuses to run if the APK has been re-signed (per branch: distinct digest pins for Main vs custom-rom)
• Disclaimers rewritten — third-party Skill and MCP disclaimers now reference Box, not Google
• Custom-rom build variant identifier fix

🛠 Under the Hood

• LiteRT upgraded to 2.1.5 (standalone artifact — no Google Play Services dependency, keeps custom-rom builds working on GrapheneOS / LineageOS / CalyxOS)
• LiteRT-LM upgraded to 0.11.0 → 0.12.0
• New small UI helpers: ScrollToBottomButton, SteadinessMonitor for image capture stability
• SystemPromptHelper + SystemPromptRepository foundations (full per-task system-prompt persistence coming in a future release)
• Interface bump: LlmModelHelper.initialize now accepts a taskId, resetConversation accepts optional initialMessages — invisible to users, eases future merges

⚠️ Known Issues / Deferred

• MCP in regular AI Chat: temporarily removed pending a proper redesign. MCP works in Agent today; AI Chat will get it back once the engine/UI plumbing is rebuilt properly.
• LiteRT-LM setPerformanceMode not yet exposed in the Kotlin SDK — sustained-performance hints are limited to the OS window-level call. Pixel 10 Pro Fold (Tensor G5) GPU decode-throttle behavior is unchanged from v1.0.10.
• Conversation history isolation between Agent and AI Chat may bleed in edge cases — fix coming in a future release.

📦 Compatibility

• Min Android version: 16+

• Recommended: Pixel 6a or newer, Snapdragon 8 Gen 2 or newer, Tensor G3 or newer

  APK Variants

Box_v1.0.11_Main_Signed_Release.apk - Stock Android (Pixel, Samsung, etc.) — includes Gemini Nano and Background Removal

Box_v1.0.11_custom-rom-support_Signed_Release.apk - GrapheneOS, LineageOS, CalyxOS — no AICore features

What's new

• Audit a file or log locally. Point Box at a file — say, a system log — and ask it to audit it (e.g. "audit /var/log/dmesg for anything suspicious"). It reads the whole thing, even files far larger than the model can normally take in one go, and writes a summary report. A progress bar shows how it's going, and you can stop it any time.
• On-the-fly file access. The assistant is no longer limited to a single workspace folder. Turn on Allow access outside the workspace (Preferences → Tools → Filesystem) and it can request any file — Box asks your permission for each path, showing exactly what it wants to open. Choose Allow once, for this chat, or always, and manage the paths you've always-allowed from the same settings page.

Also fixed

• The context-window setting now updates the usage indicator straight away, and adjusting it no longer kicks off repeated model reloads while you change the value.
• A file audit that's stopped partway through now shows what it actually found, instead of incorrectly reporting that nothing turned up.

Install

sudo apt install ./box_0.2.0_amd64.deb

Dependencies are pulled automatically. Launch Box from your app menu or run box. Upgrading from v0.1.0 just installs over the top — your chats, models, and settings are kept.

Requirements

• Ubuntu/Debian (amd64) with a GTK4 / libadwaita desktop session
• Python 3.14 (pulled as a dependency)
• ~3–4 GB free storage for a model
• CPU-only works fine; GPU acceleration is faster but optional

Checksum

Box for Linux v0.1.0

The first release of Box for Linux — a native GTK4 / libadwaita desktop app that runs AI models fully locally on your machine. Chat, real-time voice conversation, live camera vision, document Q&A, and web/file tools — all offline, no account, no telemetry.

Highlights

• Local chat with .litertlm models (Gemma 4 E2B / E4B) — streaming output, Markdown + LaTeX rendering, multimodal attachments
• Voice & conversation — voice messages, hands-free voice conversation mode with push-to-talk, offline Piper TTS in six voices
• Live camera vision — point a webcam and ask; one-shot capture or continuous Vision Mode
• Knowledge Base (RAG) — index PDFs, docs, and images per-chat or into reusable Notebooks; answers cite their sources
• Tools & agent mode — HTTPS-only web search and a workspace-scoped filesystem tool, with multi-step agent chaining
• Persistent memory — save a fact once, recalled across all chats
• Six themes — Catppuccin (Mocha / Latte / Frappé / Macchiato) and Dracula (classic + Pro)
• Everything off by default — every capability is a separate opt-in switch, with permission prompts before any tool touches your machine

Install

sudo apt install ./box_0.1.0_amd64.deb

Dependencies are pulled automatically. Launch Box from your app menu or run box. On first run, Box offers to download a model (Gemma 4 E2B, ~2.59 GB) — downloaded once, then used entirely offline.

Requirements

• Ubuntu/Debian (amd64) with a GTK4 / libadwaita desktop session
• Python 3.14 (pulled as a dependency)
• ~3–4 GB free storage for a model
• CPU-only works fine; GPU acceleration is faster but optional

Checksum

Box v1.0.10

Gemini Nano Hub — Main branch only

A new "Nano" task in the LLM section gives Pixel 9+ users access to six on-device ML Kit features powered by Gemini Nano via AICore, Over kill the Application could already do this, however this uses NPU/TPU and works a little different.

(NPU/TPU-accelerated, fully offline):

• Summarize — condense long text into a concise summary
• Proofread — grammar and style corrections with diff view
• Rewrite — rewrite text in a chosen tone (Formal, Casual, Shorter, etc.)
• Chat — multi-turn conversation with Gemini Nano (see below)
• Describe Image — caption any photo from your gallery or live camera
• Speech → Text — on-device speech recognition

Proofread and Rewrite support document attachment — pick a PDF, TXT, or MD file and its content is passed as context.

▎ First use: AICore will download Gemini Nano in the background (~1–2 GB). Wi-Fi recommended. One-time only.

💬 Nano Chat — Persistent multi-session

Nano Chat now saves conversation history to the existing encrypted SQLCipher database. Each session is listed by name, resumable, and can be
renamed or deleted. Sessions are auto-titled from the first message. Long-press any bubble to copy it.

Document Q&A

📸 Describe Image — Live Camera

Describe Image now has a Gallery tab and a Live Camera tab. Tap Capture to snap the current camera frame and send it to Nano for a description — no
need to save the photo first.

Themes — Catppuccin & Dracula

Three-way theme picker in Settings: System (Material You) / Catppuccin (14 accent colours) / Dracula (7 accent colours). Your choice persists
across restarts with no first-frame flicker.

Background Removal — Main branch only

New tool in the Image Generation section. Removes the background from any photo using ML Kit Subject Segmentation — outputs a transparency-preserving PNG. Includes a "Trim transparent edges" option. Save or share the result.

Security

Two new toggles in Settings:

• Tap jacking protection (on by default) — blocks touch events when an overlay is detected on the window
• Accessibility data sensitivity (off by default) — hides app content from untrusted accessibility services. Note: incompatible with TalkBack if
  enabled.

Bug Fixes

• Snapdragon 8 Elite NPU model — download progress was showing >100% (e.g. 112%) because the declared file size was copied from the GPU model in
  error. Corrected to 3.02 GB. NPU-only accelerator changed to npu,gpu,cpu so a failed NPU init falls back to GPU instead of crashing the app.
• Startup crash / Retry crash — fixed a null-pointer crash in the model list sort comparator that triggered on first launch and again when tapping
  Retry on the error dialog.
• LaTeX in table cells — inline math expressions inside markdown table cells no longer wrap across multiple lines.
• Import button — removed the "GGUF · LiteRT" subtitle from the home screen import button.

APK Variants

Box_v1.0.10_Main_Signed_Release.apk - Stock Android (Pixel, Samsung, etc.) — includes Gemini Nano and Background Removal

Box_v1.0.10_custom-rom-support_Signed_Release.apk - GrapheneOS, LineageOS, CalyxOS — no AICore features

MCP Server feature will be in the next release !