Box v3.3.3

Faster on-device AI, a new image generator, 15 new models, and Box in French and Portuguese.

This release rebuilds Box on the latest versions of its core AI engines — and the result is simply faster.
GGUF chat and image generation both improved noticeably on the same hardware, with no settings to
change. On top of that: a brand-new powerful image generator model, a big batch of new assorted models, and the first
translations merged from upstream.

🎨 Bonsai Image 4B — a new, sharper image generator

A new text-to-image model, and now the recommended choice for image generation in Box.

• 512×512 output — double the resolution of the FLUX.2 klein tile
• Smaller download than klein or Z-Image: ~4.3 GB, versus 7.4 GB and 10.6 GB
• Runs entirely on your phone, 4 steps, no internet at any point
• A couple of minutes per image

Bonsai is a ternary-weight build of FLUX.2 [klein], and the quality genuinely surprised us for
something running on a phone. It runs on the processor rather than the graphics chip, so expect a
wait — but the results are worth it.

Tip: run the output through EDSR ×4 (Upscale) for a 2048×2048 image.

A note on image generation: FLUX.2 klein, Z-Image and Bonsai all produce better results than the
bundled Stable Diffusion models.

⚡ Faster chat and faster image generation

Box's AI engines have been updated to their latest versions — months of upstream improvements in one
go. You don't need to do anything; it's just quicker.

• GGUF chat is noticeably faster on the same phone and the same model
• Image generation is faster too
• Whisper transcription and Gemma/LiteRT chat behave exactly as before

🌍 Box speaks French and Portuguese

Box's own interface can now be switched to another language, under Settings → Language:

• Français
• Português and Português (Brasil)

Translation is a work in progress (we're following upstream on this one): parts of the app are
translated and the rest stays in English for now, so you'll see a mix. It's not much at the moment,
but it's a start. More languages and fuller coverage to come.

🧠 15 new models

Ask about a photo — four new vision models. Show Box an image and ask what's in it, what a sign
says, or what's happening:

• SmolVLM2-2.2B (1.5 GB) and SmolVLM2-500M (just 0.36 GB — one of the smallest that works at all)
• InternVL3.5-2B (1.6 GB) and InternVL3-2B (1.4 GB) — good at reading text in pictures

New chat and reasoning models:

• Qwen3.5-0.8B — stays fast as the conversation grows, thanks to a memory that doesn't balloon with every message
• Phi-4-mini-reasoning — Microsoft's maths and step-by-step problem solver
• Polaris-4B Preview, Nanbeige 4.2 3B, SmolLM3-3B, Jan-nano — reasoning models that lay out their thinking before answering
• Ministral 3 3B in both Instruct (answers straight away) and Reasoning flavours
• OLMo-2-1B Instruct — fully open weights, data and training code
• Granite-4.0-H-1B — an unusual IBM model whose memory stays a fixed size however long you chat
• LFM2.5-1.2B-JP — tuned specifically for natural Japanese conversation

Gemma 4 12B is now a 560 MB smaller download for exactly the same capability.

✍️ Model descriptions rewritten in plain English

Nearly 70 model cards have been rewritten. Previously many read like release notes for engineers —
"density-map crowd-counting model", "192-d embedding", "hybrid Mamba2 + attention". Now each one
opens by telling you what it actually does and when to pick it, with the technical specifications kept
at the end for those who want them.

The chip-specific builds are also much clearer: cards now say "For Pixel 10 only" or
"For Snapdragon 8 Elite phones only" up front, so it's obvious which download suits your phone.

🔒 Fixes

• Chat now works properly on de-Googled Android (custom-rom-support build). On GrapheneOS,
  AOSP, crDroid, LineageOS and similar, AI Chat was silently falling back to the processor no matter which
  accelerator you chose — and the Tensor G5 model wouldn't load at all, failing with an
  "Input tensor not found" error, even though the Benchmark screen ran the very same model on the
  TPU. Both are fixed: chat now uses your GPU or TPU as intended.
• Large downloads resume by themselves. A stalled download of a very large model (Gemma 26B,
  Z-Image) used to stay stuck until you closed and reopened Box. It now retries automatically and
  picks up where it left off.

🔧 Under the hood

• Updated to the latest AI engine versions across the board
• 28 library updates, including a newer build toolchain
• Image generation memory use may be slightly higher than before on some devices and some models

Known limitations

• Bonsai runs on the processor, not the graphics chip — its diffusion model is too large for the
  GPU path — so allow a couple of minutes per image and around 8 GB of RAM.
• App translations are partial. Untranslated text stays in English.
• Voice input needs a text-to-speech engine installed. On stock Android this is present by
  default; on de-Googled ROMs, choose a voice in Settings first.

🙏 Thanks

Huge shoutout and thanks to all the great and wonderful people making these models available to
everyone — Box would not be possible without the hard work of the community!

If you'd like to see your model, or a model you know of, featured in Box, send me an email or open a
GitHub issue. Box prioritises the LiteRT backend, so LiteRT-LM models are preferred for several
reasons — but I'm open to quality GGUF suggestions too!