User repository on GitHub
Jev-style parallel constrained decisions for any MLX model on Apple Silicon. Typed, schema-valid JSON in one forward pass.
- Forks
- 8
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Repos & SDKs
Jevmlx
An Apple Silicon decision layer that scores every allowed option of a schema (booleans, enums, multi-selects) from MLX model logits in one prefill, assembles schema-valid JSON itself, and exposes a System One-compatible endpoint.
- Stars
- 70
- Language
- Python
- License
- MIT
- Reading time
- 4 min
- Published by
- Community
- Author
- bnsd55
- Last verified
- Sep 2026
About this project
jevmlx turns a schema of fields (booleans, enums, multi-selects) and a context string into a single batched forward pass on a local Apple Silicon model. Every allowed option for every field is scored from logits in one prefill — no text generation — and the JSON is assembled from the winners, with a probability per field. Requires an Apple Silicon Mac (M1+) and Python 3.12+. First use of a model alias downloads weights (2 GB fast, 4.5 GB quality).
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- One prefill reads a logits vector per field with a restricted softmax over allowed options; no text is generated and the JSON is valid by construction.
- Model aliases resolve to MLX 4-bit Qwen2.5 builds: quality (7B, default), fast (3B), and test (1.5B); first use downloads roughly 2-4.5 GB.
- Ships a CLI, an HTTP server with a /v1/systemone route, a zero-dependency TypeScript client, and golden-prompt plus benchmark tooling.
- Ordered enums add ordinal telemetry (expected_index, expected_score_normalized) with no extra model call; allow_none_of_above adds an explicit escape option.
- FieldResult includes probability, alternatives, probability_margin, and legal_mass as a leakage signal.
Quickstart
bash
git clone https://github.com/bnsd55/jevmlx && cd jevmlx && ./setup.sh
jevmlx decide --preset support_triage --json
jevmlx serve --model fast --port 8000Watch out
MIT-licensed and not affiliated with TypeSafe AI. Requires an Apple Silicon Mac (M1+) and Python 3.12+, downloads weights on first use, and its probabilities are over the supplied options, so thresholds need fitting on labeled data.
Questions about this entry
Short answers drawn from the linked project or page.
No text is ever generated. Every allowed option for every field is scored from one logits vector per field with a restricted softmax over the options you listed, and the JSON is assembled in code, so it is valid by construction instead of parsed and retried. It is a scoring tool though: free-form output such as a summary still needs a generation model.
A restricted softmax puts all probability mass on the listed options, so overlapping options starve each other. Split them into separate boolean fields or use a multi-select. It also cannot say none of these unless you add an escape choice or set allow_none_of_above, which requires an Optional field.
Apple Silicon only. It requires an M1 or later Mac and Python 3.12+, and the aliases resolve to MLX 4-bit Qwen2.5 builds, so first use downloads roughly 2 GB for fast and 4.5 GB for quality. It can send the same prompts to an OpenAI-compatible server instead, but that costs one request per field and only exposes the server's top-k logprobs.
Both, and the README separates them. TypeSafe rows are cited as official; local rows are marked as measured by contributors on the public examples. The benchmark doc covers methodology against the public consensus, a majority baseline, batched-versus-single parity and a drift band between runs. It is not a controlled comparison against Jev on your own data.
jevmlx serve loads the model once and runs one serial worker on the Metal GPU behind a bounded queue. It serves POST /decide, POST /v1/systemone, GET /v1/models, GET /health which is always 200, and GET /ready which is 503 until load and warm-up finish. A full queue returns 429 with Retry-After and an oversized request returns 413, both before model work. A zero-dependency TypeScript client mirrors the JSON shapes.
Fit the cut on labeled data rather than by feel: abstain_below_margin=X withholds any field whose probability_margin, top one minus top two, falls below it, so the abstention rate matches your review capacity. legal_mass reports how much probability mass stayed inside the allowed continuations, and a low value is a leakage signal. calibration fits multi-selection and prior_correction subtracts the neutral-context prior.
There is something to start from: decide --preset loads a bundled schema and context, including support_triage with 30 fields including ordered frustration_level and churn_risk, code_security and fintech_fraud with 28 each, content_moderation with 20, inbound_email with 19, and high_cardinality_255 with 4 for latency scaling. You can also pass a Pydantic model or a JSON schema, and decide_many handles batches of contexts.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
An educational Jev-like inference project for Apple Silicon: Qwen3.5-0.8B scores candidate answers from shared multimodal context instead of generating text, with browser, CLI, and HTTP interfaces.
Jev Repos & SDKs#community#python#open-models
Community
hr98w/jev-visualPython
hr98w/jev-visualPythonOpen multilingual System 1 decision models with published checkpoints for choice, score and noul questions, plus a router that dispatches each request to the right checkpoint in one forward pass.
Jev Repos & SDKs#community#python#open-models
Community
NandhaKishorM/layaPython
NandhaKishorM/layaPythonNokia Applied Research's package that turns any open LLM into a Jev-style decision model: typed questions are read from a single prefill, and the calibration fixes need no fine-tuning.
Jev Repos & SDKs#community#python#open-models
Community
nokia-applied-research/AnyJevPython
nokia-applied-research/AnyJevPythonTurns local text and vision models into Jev-style decision models by reading probabilities from prefill logits alone, with SGLang, Transformers, and Apple Silicon MLX backends.
Jev Repos & SDKs#community#python#open-models
Community
Yinsongxu/LLM2JevPython
Yinsongxu/LLM2JevPythonTogether AI's open recipe and weights for a Jev-inspired decision model fine-tuned on Qwen3.5-4B, with the full data pipeline, training config, and saved benchmark results.
Jev Repos & SDKs#community#python#open-models
Community
togethercomputer/tev1Python
togethercomputer/tev1PythonAn open 0.6B System One decision model that reads unstructured state and typed questions in one forward pass, returning calibrated distributions with zero decoded tokens; weights are published on Hugging Face.
Jev Repos & SDKs#community#python#open-models
Community
malevrigns/agent-jevPython
malevrigns/agent-jevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.