BOAR v2.0.0-rc1

Oct 22, 2026.

BOAR v2.0.0-rc1 is the native Android app, written in Kotlin on llama.cpp. It replaces the Expo app (v1.x). It is a release candidate: please test it and tell us what breaks.

What's new

• A native app. Same package (team.sopa.aoair), same signing key. It installs as an update over v1.0.0 and v1.1.1 and keeps your downloaded models.
• Answers with sources, offline. Knowledge packs hold the library BOAR searches. The Wikipedia Vital Articles pack has about 50,000 articles in 164 MB. Each answer shows the articles it used, so you can check them.
• Personal Files. Import your notes, PDFs, spreadsheets (.txt, .md, .csv, .json, .pdf) in Knowledge → Personal Files, and ask about them with the Personal Files style. After an import, "Ask about them" opens a chat in that style. Other file types are refused with a clear message.
• A tour while setup downloads. A "While you wait" card opens a short tour of what BOAR can do, with the download live on top. It's also in Settings → System Recovery → See the tour.
• Deep Research for longer, multi-step answers, and answer styles (Succinct & Direct, Thorough & Detailed, and Personal Files, which answers only from your own documents and says so when they don't have the answer).
• PIN lock. Chats are sealed by the PIN. With a PIN set, BOAR's screen is blank in recent apps and in screenshots.
• "BOAR is answering…" An answer keeps going when you leave the app. BOAR asks once, with your first question, to show this notification; saying no is fine.
• Memory freed when you leave BOAR (Auto, on by default). The next answer takes a few seconds longer to start. Close BOAR in the menu stops any answer, frees the model and closes the app.
• Hugging Face model discovery. Search, add and download GGUF models. It needs internet only while it searches and downloads. Models that need a Hugging Face account are marked: BOAR downloads without an account, so it can't get them.
• Fixed first-aid cards for a few questions where small models gave unsafe answers: snakebite, box jellyfish and frostbite.
• 16 KB memory pages. Every native library is 16 KB-aligned, so BOAR runs on the newer phones that use 16 KB memory pages (Android 15 and newer) with no compatibility warning.
• Deleting a chat asks first. Settings → Memory → "Ask before deleting a chat" (on by default) turns it off if you prefer one tap.
• Benchmark runner over adb, for anyone who wants to measure models on their own phone.
• No results sharing. Evaluation sharing is off in this build: there is no Share button, and runs stay as files on the phone.
• Boar Labs, on by default, with five quick-answer labs and Boar Vitals. Facts now also knows each country's driving side, calling code, mains frequency and emergency numbers (195 countries). See below.

Answers come from the phone. Nothing leaves it unless you download a model or a pack.

Install

From v1.x: install the APK over your current BOAR. It updates in place and keeps your models. You don't need to uninstall.

Fresh install: on first launch, setup downloads a model, the search model and the Vital Articles pack once (about 1.3 to 2.7 GB, depending on the model). After that, BOAR works offline.

BOAR suggests models by the phone's memory:

Phone memory	Recommended in setup
10 GB and up (e.g. 12 GB phones)	Qwen3 4B, "Recommended"
10 GB and up	LFM2.5 8B, "Fast" (quick, but check what it says)
Under 10 GB (e.g. 8 GB phones)	Qwen3 1.7B, "Recommended · light"
Any	Qwen2.5 1.5B, Gemma 4 E4B and Qwen2.5 7B are listed too
Optional, with Big Boar	Qwen3.6 35B (12.3 GB download)

Built locally. This APK was built on our machine from the tag v2.0.0-rc1, not by a CI service, and signed with BOAR's key:

Signer: CN=BOAR, O=sopa.team
Certificate SHA-256: 7cf514f9ab8253cbd78eb7d112d66fd05c2ad27e6fc987bf8f3276a7eedf1f01

Check it with apksigner verify --print-certs boar-v2.0.0-rc1-arm64.apk.

What's in Boar Labs

Labs are experimental features, labelled beta (tested more) or alpha. Settings → Experimental holds Boar Labs. Labs don't slow the answers; the ones that load a model (Designer, Big Boar, Wisp) use extra memory while they run. If a lab misbehaves, switch it off there and open an issue.

Lab	What it does	At first launch
Calculator	Exact answers to everyday calculations (battery, fuel, hiking, currency, water, temperature), without the model	On
Facts	One fact about a country, city, mountain, river, landmark or award, from a Wikidata table	On
Cohazard	A fixed carbon monoxide safety answer when you plan to run a fuel-burning device indoors	On
Distance	How far apart two cities are, and an estimate of the time to walk, cycle or drive it	On
Polyglot	Says so at once when you ask in a language BOAR doesn't speak yet	On
Guard	Checks a yes-or-no verdict in a model's answer against the rest of the answer before it's saved	Off
Designer	Draws an illustration from a few words, on the phone (needs a model download)	Off
Press	Collects answers into a basket and compiles them into a PDF or Markdown file	Off
Big Boar	Runs a 35B mixture-of-experts model, reading its experts from storage. Slow	Off
Wisp	A small mixture-of-experts model (OLMoE-1B-7B) on a tiny engine that streams from storage	Off
Nomad	A compass from the phone's own sensors, no location needed	Off
Lunar	The moon tonight and this month: phase, lit share, next full and new moon	Off
Vitals	Shows the phone at work: heat, core clocks, memory and the chip	On
Signboard, Truffle Hunt, Sanctuary	Shown as locked tiles: coming later	—

Labels: Calculator, Facts, Cohazard, Distance and Polyglot are beta; Vitals is live; Wisp is under review; the rest are alpha.

Tested on

Three lab devices: Lab device 12 GB (X6), Lab device 8 GB, and Lab device Pixel 8a with GrapheneOS. On the Pixel 8a (Oct 21): install over the previous build, the 16 KB check, answers, the notification and the delete confirmation.

Speed, Succinct, 512 tokens:

Phone, date	Model	Speed	First word
X6, Oct 18	Qwen3 4B	6.8 tok/s	12.5 s
X6, Oct 18	Qwen2.5 1.5B	16.9 tok/s	4.7 s
8 GB, Oct 18	Qwen3 4B	7.7 tok/s	not measured
8 GB, Oct 18	Qwen2.5 1.5B	19.7 tok/s	4.9 s

• Big Boar, X6, Oct 19: Qwen3.6 35B answered all 73 of r4to's questions (73/73). Median speed 3.35 tok/s, about 2.5 min per answer. Battery went from 27.2 to 41.2 °C; thermal state stayed normal.
• Chaos tests: 60 random moves per run, on both phones, 0 crashes. After the Oct 20 fixes, the model is freed before the phone's system freezes BOAR in the background (Lab device 8 GB, Oct 20).
• PIN screen: recent apps and an in-app screenshot came out blank (X6, Oct 20).
• Qwen3 1.7B is new in this release; its first runs on our build are on the Lab device 8 GB during the rc1 tests.

Known limits

• Labs features are experimental. They may change or be removed.
• Adaptive routing is experimental and off by default.
• On 8 GB phones, Android can close BOAR while a big model is loaded.
• Big Boar is slow: an answer takes minutes (2 to 4 tok/s on the X6, Oct 19).
• The 35B model (12.3 GB) and Designer's model are large downloads.
• Tested on three lab devices, one of them with GrapheneOS.
• Voice input needs a speech recognizer app. De-Googled phones may not have one.
• Small models still make mistakes. Check the sources when it matters.
• Personal Files is strongest with Qwen3 4B: 17 of 18 questions about our test files right, 0 invented (Lab device 12 GB, Oct 21). Smaller models can invent an answer when the files don't hold it.
• In the default styles, a question about your own files can be answered from the library instead. Use the Personal Files style for them.
• Spreadsheets (.csv): the model can mix up columns.
• The PIN seals chats, but imported documents and the database are not encrypted yet.

Checksums

File:    boar-v2.0.0-rc1-arm64.apk
SHA-256: a79011c6eeb7fcdc33b8ef9daf12b2b3d0faae947d0bfbde9a636396dfacf8a5

Check it with sha256sum boar-v2.0.0-rc1-arm64.apk.

Offline APK (no download at all)

boar-2.0.0-rc1-offline-arm64.apk (2.06 GB) carries Qwen3 1.7B, the search model, the Vital Articles pack and Designer's drawing model inside it. On first launch BOAR copies them out and checks each one; then it works with no network at all, not even once. Same app, same signing key: it installs over BOAR v1.x or the rc1 APK. It suits phones under 10 GB of RAM, or anywhere without internet. Built from rc1 plus the offline packaging commits, on the branch feat/offline-bundle (commit 33be7bd2), with scripts/build-offline-apk.sh release. The script takes the model files from your own machine and checks each one against the catalog's SHA-256 before bundling it.

Tested on the Lab device 8 GB in airplane mode (Oct 22): install, unpack (about 3 seconds there), and an answer with sources, all offline.

File:    boar-2.0.0-rc1-offline-arm64.apk
SHA-256: bd1f3b86d5f27898bcc1805125d9b617df0f03b1e11c4e945debf75f8e2a6c39

Also on Hugging Face, next to the 4B one: BOAR-App/boar-offline (boar-v2.0.0-rc1-offline-qwen3-1.7b-arm64.apk, the same file).

Offline APK with Qwen3 4B (3.45 GB, on Hugging Face)

Too big for a GitHub release file, so it lives on Hugging Face: boar-v2.0.0-rc1-offline-qwen3-4b-arm64.apk (repo). Same contents as the offline APK above, with Qwen3 4B instead of 1.7B, the same signing key, and built from the same branc...

BOAR v1.1.1 (Evaluator & Benchmark Release)

BOAR v1.1.1 (Evaluator & Benchmark Release)

A new look for the app and secure, hardware-attested sharing of benchmark runs. This can replaces v1.1.0, which was withdrawn due to UI rendering issues on certain Android devices.

🔒 Important Privacy & Evaluation Policy:

• boar-v1.1.1-arm64.apk is the specific version designated for evaluators, and testers who explicitly want to run and share public model evaluation results.

• All other versions will remain strictly closed to enforce BOAR's core "Nothing leaves your phone" policy. Sharing is completely disabled by default, and zero chats, logs, or metrics ever leave your device.

---

Share your benchmark runs

• Signed by your phone: Each shared evaluation run from boar-v1.1.1-arm64.apk is signed with a hardware-backed key from the phone's secure chip (Android Keystore). This guarantees that public results hold verified, authentic on-device performance metrics.
• A progress screen shows each step as it executes (prepare, challenge, key, sign, send), followed by the result and export options (JSON/CSV).
• Opt-in only: Sharing operates strictly on a manual opt-in basis within this specific signed release APK; development and standard builds refuse sharing by design.

---

New look

• Campfire and Moonlight: The two core themes from BOAR's visual identity, with Campfire as default.
• New typography: Baloo 2 for titles and Lexend for reading, with a larger base size (body 16, nothing under 12). Compact / Standard / Large font settings apply across every screen.
• Cleaner screens: Removed emojis from headers and menus; sentence case instead of ALL CAPS; flat, rounder chat bubbles; pill-shaped input with enlarged buttons.
• BOAR mascot & ambient glow: Mascot integrated in the header with a warm ember glow (Campfire) or moon halo (Moonlight) behind the chat stream.
• Fresh chat prompts: Offers short suggested questions on new chats that send immediately upon tapping.
• Taller header: Clean model naming (QWEN2.5-1.5B-INSTRUCT), RAM, and disk utilization consolidated into a single menu line.

---

Also

• The share preview screen cleanly fits inside the device's safe area.
• Includes English and Portuguese preview drafts for Privacy Policy and Terms of Use.

---

Install

Download boar-v1.1.1-arm64.apk (124 MB, 64-bit ARM Android) and verify its SHA-256 checksum:

sha256sum -c boar-v1.1.1-arm64.apk.sha256

Installs directly over v1.0.0 and v1.1.0 while preserving all downloaded models. First launch on a new phone downloads Qwen2.5-1.5B and a small embedding model (~1 GB total); after initial setup, BOAR operates 100% offline.

BOAR v1.0.0

Offline AI research app for Android (arm64). Install the APK, open it once online to download the ~1 GB default model, then it works fully offline. Verify: sha256sum -c boar-v1.0.0-arm64.apk.sha256

Knowledge pack v1

📚 Knowledge Pack v1

Wikipedia Vital Articles (Level 5) Pre-indexed offline encyclopedia pack containing article introductions across science, history, geography, and world knowledge.

• Scope: 49,832 article lead summaries.
• File Size: 164 MB (wiki-vital5.sqlite).
• SHA-256: d3b87d562baba3489f6878bf99783f50d504db94c347029771e53f6d1aecc666
• Indexing: Hybrid FTS5 BM25 + 8-bit bge-small-en-v1.5 embeddings (384-dim).
• License: Text CC BY-SA 4.0 (Wikipedia).