Repository navigation
Commonplace 0.3.0
Commonplace is an offline research assistant for Android. It answers from Wikipedia and other reference text on the phone and cites its sources. The app has no INTERNET permission.
Changes since 0.2.4
- Gemma 4 E2B on the Tensor chip. On a Pixel 10, the new Gemma pack (3.1 GB) runs Gemma 4 E2B on the Tensor G5 NPU. The first word comes after about 2 s and the full answer after 6–8 s, against 17–29 s with Ling on the CPU. With the pack installed, Gemma is the answer model; Settings can still choose Ling. The Library lists the pack only on phones with a Tensor G5 and marks it "For this phone", and Get started on such a phone lists it next to the starter set.
- Speculative decoding for Gemma on the NPU, on by default. It writes about 1.5 times faster (27–31 tok/s). Settings can switch it off.
- Think with Gemma. The Think button and "Think it through" work with Gemma, and the answer shows "Thought for N s".
- Defaults for each model. The evidence budget in Settings shows the loaded model's default, and "Use the model's default" removes your own value. Gemma keeps its full evidence on a warm phone, because the NPU does not slow down.
- One model list in Settings. Gemma and the CPU models are choices in one list. An engine choice saved by an older version no longer keeps a phone with the Gemma pack on the CPU model.
- Safe start. If Gemma fails or stops the app while it loads, the app answers with the CPU model, and Settings says why.
- LiteRT-LM 0.18.1, with a Tensor dispatch library built from the LiteRT commit that it pins. Google's prebuilt library crashes with this LiteRT-LM.
Install 0.3.0 over any earlier version. The installed packs stay.
Install
- Download
commonplace-0.3.0.apkbelow on an Android phone (64-bit ARM, Android 12 or later) and install it. - Open Commonplace and tap Open Library. Under Get more, tap the starter set marked For this phone and tap Download. The browser saves
commonplace-starter.tar(11.3 GB, 12 GB phones) orcommonplace-starter-small.tar(8.1 GB, 8 GB phones) from Hugging Face. On a Pixel 10, also download Gemma 4 E2B (Tensor G5) (3.1 GB). - Tap Install from Downloads and select the files. The app checks every byte with SHA-256 during the import.
- Ask a question.
You need 8 GB of RAM and free storage of about twice the starter set during the first import. docs/INSTALL.md lists the other packs and the install from a computer.
Packs
The Hugging Face dataset 0xPD33/commonplace-packs, folder packs-2026-09, has one new file: gemma4-e2b-npu.tar.
| File | Size | Contents |
|---|---|---|
commonplace-starter.tar |
11.3 GB | Starter set for 12 GB phones: Wikipedia starter (2M articles), Ling-3.0-tiny, Wikidata facts |
commonplace-starter-small.tar |
8.1 GB | Starter set for 8 GB phones: Wikipedia starter, MiniCPM5-2B, Wikidata facts |
gemma4-e2b-npu.tar |
3.1 GB | Gemma 4 E2B for the Tensor G5 NPU: the fastest answers on a Pixel 10 |
commonplace-breadth.tar |
2.0 GB | 13 packs: textbooks, Wikibooks, Wikivoyage, Wiktionary, DevDocs, medical and travel health pages and more |
enwiki.tar |
19.2 GB | All of English Wikipedia (replaces the starter Wikipedia) |
arxiv-abs.tar |
3.3 GB | arXiv abstracts |
stackexchange.tar |
2.5 GB | Stack Exchange questions and answers from 51 sites |
minicpm5-2b.tar |
1.4 GB | The small model alone, for a phone with the 12 GB starter set |
voice-en.tar |
0.3 GB | Voice input, English |
Verify the APK
- SHA-256 of
commonplace-0.3.0.apk:e48f5daa07d2d04040b0808e3c8f3f3ad34b50ece0e9b4073d1e87789dd06184 - Signing certificate SHA-256:
c17ec0b2fe898fa8eb6b172c14be053c5748965a8803bd0f021de737a65aa10b - Permissions: microphone, foreground service and notifications. No network permission.
Known limits
- Gemma runs on the NPU only on a phone with a Tensor G5 (Pixel 10). Google publishes no NPU file for other chips or for Gemma 4 E4B, so other phones answer with Ling or MiniCPM on the CPU.
- On 63 seed questions asked on a Pixel 10, Gemma scores 0.50 of Claude's score with web search, and Ling 0.53. Head to head, Ling wins 33 and Gemma 21, which is not significant. Gemma is weakest on numbers.
- In a long session the phone heats up, and Ling's first word takes longer on a hot phone. Gemma on the NPU kept its speed over 126 questions in 76 minutes (first word 2.8 s at most), but that test did not make the phone hot.
- Real speech recognition is tested on the desktop, not yet on a phone's microphone.