Organization repository on GitHub
Local multimodal decisions on your Mac. Jev-compatible typed probabilities with Qwen, llama.cpp and Metal.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Practices & Patterns
Openjev Multimodal
A local, open-source System One API compatible with Jev, served by Qwen models over llama.cpp and Metal: Noul, Choice and Score answers with full probability distributions and native image input, with no paid inference API in the loop.
- Stars
- 3
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- jev-skills
- Last verified
- Oct 2026
About this project
Documentation · 中文文档 · 中文 README · Benchmarks. Text and images → typed decisions. One output token per question. Runs on your Mac. A local, open-source Jev-compatible SystemOne API powered by Qwen and llama.cpp / Metal. Route messages, interpret screenshots, score rubrics or choose the next action with complete probability distributions. Three primitives: Noul (yes probability), Choice (selection + distribution), Score (expected rubric index).
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- One output token per question, with up to 255 Choice options backed by verified single-token labels and complete probability checks.
- Four profiles from a 0.8B to a 27B Qwen checkpoint, with the two largest shipping tens of gigabytes of weights plus a projector.
- Requires Apple Silicon, Python 3.11 and llama.cpp b9670 or newer built with post-sampling probabilities.
- Binds localhost by default, has optional Bearer auth on /v1/*, exposes health and limit endpoints, and fetches no remote or request-named files.
Watch out
This is an independent implementation, not affiliated with TypeSafe and not using its proprietary Jev weights, and the probabilities are conditioned on the options you supply rather than calibrated correctness guarantees. The larger profiles are heavy: quality is about 23.3 GB and max 17.6 GB of weights, each plus a 0.9 GB projector. The published measurements come from an M3 Max with 128 GB of memory and minimums for smaller machines were never benchmarked. Audio and native video are unsupported, sampled frames have to be supplied as images instead. The benchmark is a nine-task exploratory subset at 20 samples each, not a leaderboard result, and the GitHub Pages site hosts static documentation rather than the inference API. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
A different model with a compatible wire shape. It is powered by Qwen checkpoints served through llama.cpp and Metal, described as an independent implementation not affiliated with TypeSafe and not using its proprietary Jev weights. The compatibility claim is about the SystemOne request and response format: one synthetic online Jev request was used to verify the common wire shape, and that request is not part of any benchmark.
Not a leaderboard, and the README does not present it as one. It is an exploratory subset of nine measured tasks at 20 samples each, 180 displayed evaluations, with only this project's own API plotted and Wilson intervals showing sampling uncertainty. GSM8K uses synthetic numeric distractors and chess measures legal-move selection, a larger run was stopped to limit load with 886 completed receipts archived, and the chart takes the first 20 in each task's random order with no score-based filtering.
It needs Apple Silicon, Python 3.11 or newer and llama.cpp b9670 or newer built with post-sampling probabilities. The balanced default is a 4B Q4_K_M checkpoint, which is the everyday text and vision profile, while quality and max are 23.3 GB and 17.6 GB plus a 0.9 GB projector. Defaults are one inference slot and four CPU threads, and the published latency figures were measured on an M3 Max with 128 GB, so smaller-machine minimums are unknown.
Not through this service. Inference is local, state and images stay on the machine, and no paid inference API is required. It binds localhost by default, OPENJEV_API_KEY turns on Bearer authentication for /v1/*, and the collector neither fetches remote images nor reads files a request names. The one thing to keep straight is that the GitHub Pages site serves the static documentation, not the API, which runs locally.
Apple Silicon, Python 3.11+ and llama.cpp b9670 or newer. Run `brew install uv llama.cpp`, clone the repo, then `uv sync --frozen` and `uv run openjev serve`. Open http://localhost:8000/playground for live text and image input or /docs for the API; the first run downloads pinned weights.
Four: fast (Qwen3.5-0.8B, Q4_K_M), balanced (Qwen3.5-4B, the default), quality (Qwen3.6-35B-A3B) and max (Qwen3.8-27B). Quality weights are about 23.3 GB and max 17.6 GB, each plus a 0.9 GB projector, and `serve --profile quality` runs one at a time.
MIT for the original code, with model weights and datasets keeping their own licenses, detailed in NOTICE. It is an independent implementation, not affiliated with TypeSafe, and it does not use TypeSafe's proprietary Jev weights.
More Jev resources like this
38GitHub stars
Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Argos1111/jev_localPython
Argos1111/jev_localPython24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePython39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
mithalouni/system-one-openPython
mithalouni/system-one-openPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.