Is a model fine-tuned to imitate Jev, or are logits read off any local model?

Logits are read directly. You point it at a Hugging Face-compatible local causal LM and it computes probabilities during prefill, with no token-by-token decoding and no answer parsing. Nothing is trained to imitate Jev; the project states it is independent and not affiliated with or endorsed by TypeSafe.

Are those 48 ms JevBench numbers a like-for-like comparison with Jev?

Treat it as a project-local measurement. The README reports 76.2% on 231 public JevBench items with Qwen3.5-4B, P50 48 ms and P95 346 ms, against recorded Jev 1.13.0 values of P50 652 ms and P95 722 ms. Different weights sit underneath, and the details and reproduction commands live in a separate JevBench document.

Will this run on an Apple Silicon laptop, or do I need an NVIDIA box?

Three backends: SGLang, Transformers, and MLX on Apple Silicon, all through the same binary scoring interface or the HTTP service. The quick start, though, is written for Linux with a supported NVIDIA GPU, and the MLX path prefetches shared prefixes before scoring candidate suffixes.

Can I swap it in where I already call the hosted Jev API?

Yes. It exposes a compatible POST /v1/systemone endpoint, plus an offline Python API, so a client written against the System One shape can point at it instead. The usage guide is where the staged versus all submission modes and the provider setup are written down.

What do I need to run it on Linux?

Linux with a supported NVIDIA GPU, Python 3.12 or newer, and uv sync --extra sglang, then python examples/sglang_inference.py --model-path pointing at a local Hugging Face causal model; the example submits Choice, Score, and Noul questions.

What is staged candidate submission?

Staging first submits one real criteria candidate to establish the SGLang prefix cache, then submits the others so they reuse it; candidates share the state, and candidates for the same question share its instructions.

Does it accept images?

Yes. Multimodal inputs combine text and images in state or instructions, supported on SGLang, Transformers, and MLX-VLM; the Valen Sokoban demo evaluates image-conditioned single-step actions.

Jev directory

CommunityJev Repos & SDKs

LLM2Jev — Yinsongxu

Turns 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.

See highlights Open github.com

Stars: 449
Language: Python
License: APACHE-2.0
Reading time: 3 min
Published by: Community
Author: Yinsongxu
Last verified: Sep 2026

Yinsongxu/LLM2Jev

User repository on GitHub

View on GitHub

Turn local language models into Jev-style structured decision models. Get results from text and images with prefill alone—no token-by-token decoding required.

Forks: 50
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Turn local language models into Jev-style structured decision models. Get results from text and images with prefill alone—no token-by-token decoding required. October 8 - Valen Sokoban evaluation: added a visual Sokoban demo and single-step action evaluation based on Valen. September 26 - JevBench evaluation: LLM2Jev accuracy and latency on 231 public items; P50 latency is below one tenth of the Jev official result.

Quoted from the project's README. Read the full README.

Tagscommunity python open-models sglang multimodal

Works forJev for Evaluation Jev for Scoring

Used inJev for Software & DevTools

Highlights

Scores options from prefill logits only: no token-by-token decoding and no answer parsing.
Backends cover SGLang on NVIDIA GPUs, Transformers, and MLX on Apple Silicon.
Staged candidate submission reuses SGLang's Radix Cache, including a cold first request.
Exposes a compatible POST /v1/systemone endpoint for drop-in client use.
Benchmarks Qwen3-1.7B on an RTX 5090 across staged and all-at-once scoring with cold and warm caches.

Quickstart

bash

uv sync --extra sglang
source .venv/bin/activate
python examples/sglang_inference.py --model-path /path/to/model

Watch out

The SGLang quickstart assumes Linux and a supported NVIDIA GPU, and the project is independent and not affiliated with TypeSafe.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

LinkedIn: What multimodal decision models actually change, and what they do not.

LinkedInJevDirectory.org

X: Standard One SH (StandardThinking)

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 News

Hacker 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 News

Other listings of this project

Forks and copies of this project in the directory. This entry is the canonical one.

llm2jev

Turns any chat model into a Jev-compatible decision service: state, question and all options go in one prefill, and the logprobs of the option labels at a single position become probabilities. Runs on SGLang, vLLM, transformers or MLX.

Jev Tools & CLIs#community#awesomejev#catalog-import

Community

tic-top/llm2jevPython

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Multimodal Decision Models: What Clef and d1 Change

Oct 12, 2026

Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.

Jev Models & Reimplementations#multimodal#open-models#decision-model

Read article

The TypeScript SDK at the edge: Workers AI, Vercel, and validation

Nov 2, 2026

Workers AI, Vercel, and validation: how the TypeScript SDK behaves at the edge, what it returns, and where to put the schema check.

Jev Repos & SDKs#sdk#typescript#edge

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

LLM2Jev — Yinsongxu

About this project

Highlights

Quickstart

Questions about this entry

Is a model fine-tuned to imitate Jev, or are logits read off any local model?

Are those 48 ms JevBench numbers a like-for-like comparison with Jev?

Will this run on an Apple Silicon laptop, or do I need an NVIDIA box?

Can I swap it in where I already call the hosted Jev API?

What do I need to run it on Linux?

What is staged candidate submission?

Does it accept images?

Reactions & coverage

Related terms

Other listings of this project

More Jev resources like this

From the Jev guides

From the Jev AI model community