PocketJev
Experimental iPhone app for on-device visual decisions: Qwen3-VL on MLX reads option-letter logits for 2-26 choices about a photo.
Overview
PocketJev is an experimental iOS app that turns a small vision-language model into a fast multiple-choice decision tool on the iPhone. You present a question and 2 to 26 choices about a camera capture or photo, and it reads the next-token logits for the option letters and shows the model's relative score for each choice, with no free-form text generated. Uses include yes/no/unknown visual checks, object or part presence, framing checks and pass/retry/rebuild review flows.
It runs mlx-community/Qwen3-VL-2B-Instruct-4bit (about 1.78 GB, Apache-2.0) on device after a first download, with no image upload or telemetry path, and decisions take about a second per the list summary. Features include editable questions, reusable option sets, a continuous mode every 1, 2 or 5 seconds without queuing stale frames, haptics and a Japanese UI. The README is explicit that displayed percentages are relative scores among the supplied choices, not calibrated probabilities.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Is this a cat? | YES · NO · UNKNOWN |
| Choice | Review outcome | PASS · RETRY · REBUILD |
How it works
- Capture a photo or use the live camera.
- Enter a question and option set.
- Read option-letter logits from Qwen3-VL on MLX.
- Display relative scores and the selected option.
Strengths
- Fully on-device; no image uploads.
- Custom option sets.
- Continuous mode for live checks.
Caveats
- Scores are not calibrated probabilities.
- Requires a 1.78 GB model download.
At a glance
- Type
- Model
- Status
- prototype
- Language
- Swift
- Runtime
- iOS app, MLX Swift
- Has code
- yes
Use cases and integrations
How curated lists describe it
On-device iPhone visual decisions with MLX + Qwen3-VL option logits. Camera + 3-choice, no text generation, ~1s, no photo saved.
AnotiaWang/awesome-decision-modelsOn-device iPhone visual decision tool using MLX and Qwen3-VL direct option logits.
hellogumbo/awesome-jev
Frequently asked questions
Are the percentages probabilities?
No, they are relative scores among the supplied choices.
Are photos uploaded?
No, inference runs on device with no upload path.
Which model is used?
Qwen3-VL-2B-Instruct in 4-bit MLX weights.
Related projects
llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Python library that imitates Jev-style typed decisions (choice, score, yes/no) on local LLMs via Ollama, vLLM or llama.cpp.
ProjectprototypePythonMLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonTurns any local LLM into a Jev-style engine via masked-logit softmax, or proxies OpenAI-compatible APIs or Jev; includes a chat crush radar.
ServiceprototypePythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRustEducational C11 library and CLI that turns DiffusionGemma 26B into a CPU-first OpenJev-style decision engine reading candidate logits.
ProjectresearchC