PocketJev

September 17, 2026 · View on GitHub

Pocket-sized, on-device visual decisions for iPhone.

PocketJev is an experimental iOS app that turns a small vision-language model into a fast multiple-choice decision tool. Instead of asking the model to write a long description, PocketJev presents a question and 2–26 choices, then reads the next-token logits for A / B / C / ... and shows the model's relative score for each choice.

Typical uses include:

  • YES / NO / UNKNOWN visual checks
  • object / part presence checks
  • framing and composition checks
  • PASS / RETRY / REBUILD style review flows
  • quick camera-based classification with custom option sets

PocketJev is a decision UI, not a calibrated classifier. The displayed percentages are relative scores among the supplied choices; they are not probabilities that the answer is correct.

Real-device example

PocketJev on iPhone answering NO to the question Is this a cat? for a photo of a fries carton

In this iPhone example, the question is "Is this a cat?" and the image is a fries carton. PocketJev selects NO from YES / NO / UNKNOWN. The displayed scores are still option-relative model scores, not calibrated confidence.

Features

  • Native SwiftUI iPhone app
  • Camera capture and Photos picker
  • Direct option-logit scoring; no free-form answer generation in the decision path
  • Editable questions
  • Reusable option-set dropdowns with a separate editor
  • Default YES / NO / UNKNOWN option set
  • 1 / 2 / 5 second continuous-decision mode without queuing stale frames
  • Continuous mode keeps the live camera visible while the captured frame animates into the result area; the previous confirmed result stays visible until the next decision finishes
  • Light haptic feedback after a decision
  • Japanese UI when the iPhone language is Japanese; English otherwise
  • On-device MLX inference after the model has been downloaded
  • No image-upload or telemetry path in the app

Model

PocketJev currently uses:

mlx-community/Qwen3-VL-2B-Instruct-4bit

The model is downloaded on first use and cached on the iPhone. The current Hugging Face artifact is about 1.78 GB, uses 4-bit MLX weights, and is published under the Apache-2.0 license.

The model weights are not included in this repository.

How it works

camera image / photo
        +
question + option set
        ↓
Qwen3-VL on MLX Swift
        ↓
next-token logits for A/B/C/...
        ↓
relative score per option

The app intentionally avoids generating a natural-language response for the decision itself. This keeps the output simple and avoids JSON parsing / repair.

The current iPhone implementation still re-encodes the image for every decision. Shared vision-feature / prefix reuse from the macOS prototype is not yet ported.

Requirements

  • Xcode 26 or newer is recommended for the current project configuration
  • iOS 18+
  • An iPhone with enough memory to load the 2B 4-bit VLM
  • Network access for the first model download
  • XcodeGen to regenerate the Xcode project

The underlying mlx-swift-lm package supports iOS, and this project targets iOS 18.

Build

git clone https://github.com/NullPo-jp/PocketJev.git
cd PocketJev

xcodegen generate
open PocketJev.xcodeproj

In Xcode:

  1. Select the PocketJev target.
  2. Choose your Development Team.
  3. Change the example bundle identifier (com.example.PocketJev) to a unique one.
  4. Select your iPhone and Run.

For command-line builds, SwiftPM plugin/macro validation may need to be allowed or explicitly skipped depending on your Xcode security settings.

Prebuilt IPA

GitHub Releases include a prebuilt PocketJev-vX.Y.Z.ipa plus its SHA-256 checksum. The IPA is a generic Release build intended for compatible sideloading environments (for example TrollStore where supported); it is not App Store or TestFlight signed.

  • PocketJev itself requires iOS 18+.
  • The IPA does not contain the Qwen3-VL model weights; the model is downloaded on first AI setup just like a source build.
  • The public IPA uses the stable bundle identifier io.github.nullpojp.PocketJev.
  • The app keeps the com.apple.developer.kernel.increased-memory-limit entitlement, but entitlement behavior ultimately depends on the installer / device environment.

To reproduce the IPA locally:

./scripts/package-ipa.sh 0.1.1

Output is written under dist/ and is intentionally ignored by Git.

Language behavior

PocketJev intentionally keeps localization simple:

  • iPhone language = Japanese → Japanese UI / 判別不能
  • any other iPhone language → English UI / UNKNOWN

Custom option sets are user data and are not automatically translated.

Privacy

Images are processed locally by the on-device model. The app does not contain an image-upload or telemetry path. Network access is used to obtain the model from Hugging Face on first setup. See PRIVACY.md.

Accuracy notes

This is an experimental decision interface, not a production quality-control system. A broad question such as "Is this render acceptable?" can be less reliable than smaller observation questions such as "Is any part cut off by the image border?" or "How many knobs are visible?".

If a wrong decision would matter, add an explicit UNKNOWN / 判別不能 option and verify the workflow with your own held-out examples.

Tech stack

  • Swift 6 / SwiftUI
  • AVFoundation
  • MLX Swift
  • mlx-swift-lm
  • Hugging Face Swift packages
  • Qwen3-VL 2B 4-bit MLX model

Status

Working prototype. Camera input, option-logit scoring, custom option sets, localization, model caching and continuous mode are implemented. Quantitative iPhone latency / accuracy benchmarks are not yet published.

License

PocketJev source code is available under the MIT License.

Model weights and third-party dependencies keep their own licenses. In particular, the current Qwen3-VL MLX model is distributed separately under Apache-2.0.

日本語

PocketJev は、iPhone のカメラや写真を 長文生成ではなく少数択判定 に 使う実験的なオンデバイスAIアプリです。

質問と2〜26個の選択肢を与え、Qwen3-VL が次に出そうとする A / B / C ... のlogitを直接読み、各選択肢の相対スコアを表示します。

初回だけ約1.78GBのモデルをHugging Faceから取得し、その後の推論はiPhone上で 実行します。表示される割合は正解確率として校正された値ではありません。

日本語設定のiPhoneでは日本語UI、それ以外では英語UIになります。 初期選択肢は日本語では YES / NO / 判別不能、それ以外では YES / NO / UNKNOWN です。

上の実機スクリーンショットでは、ポテトの容器に対して「これは猫ですか？」と 質問し、YES / NO / 判別不能 から NO を選んでいます。表示値はあくまで候補内の 相対スコアで、校正済みの正解確率ではありません。

実用する場合は、いきなり「この画像は合格？」と聞くより、 「画像端で見切れているか」「部品が何個見えるか」のような観測質問へ分解する 方が扱いやすい設計です。