Do I need an API key at inference time?

No. The 1.7 GB of weights download once from Hugging Face, then everything runs locally through mlx-vlm at no per-call cost; requests from web pages are refused by host and origin checks.

How does escalation work?

Every answer carries a probability for every option, so you act on confident answers and send low-confidence ones to a frontier model; the README shows a 0.8 confidence threshold example.

What hardware does intern-decision-mlx need?

An Apple silicon Mac (M1 or later) on macOS 14 or newer with Python 3.10+; pip pulls about 750 MB of mlx-vlm dependencies, and the first run downloads 1.7 GB of weights.

How close is it to the lab's reference model?

It matched the lab's PyTorch implementation on 67 of 67 fields across 30 varied requests, with a maximum calibrated probability difference of 0.023, which the README attributes to bf16 rounding.

Can I use it as a safety gate?

No. The README and the lab's card both say not to: it scores 64.48 on WildJailBreak against Jev's 96.29, so its judgement is weak even though its calibration (ECE 0.066) is good.

How do I speed up a decision?

Shrink the image. Latency grows about 1.2 ms per input token and a 1080p screenshot is about 2,300 tokens, so --max-side 1024 cuts a 3.6 s decision to 0.9 s; it is off by default so outputs match the reference exactly.

Is the server safe to expose on my network?

By default it binds localhost, handles one request at a time, refuses requests from web pages, and caps body size; passing --host 0.0.0.0 exposes it with no authentication, so only do that on a trusted host.

CommunityJev Models & Reimplementations

Intern Decision Mlx

Serves Shanghai AI Lab's Intern-Decision-0.8B vision decision model on Apple Silicon behind a /v1/systemone-shaped endpoint: screenshots plus typed questions in, calibrated Choice, Score and Noul probabilities out, for thresholding and escalation.

See highlights

Open github.com

Stars: 1
Language: Python
License: APACHE-2.0
Published by: Community
Last verified: Oct 2026

dex0shubham/intern-decision-mlx

User repository on GitHub

View on GitHub

Screenshot + questions in, calibrated typed decisions out. Intern-Decision on Apple silicon: 0.9 s per 1080p screenshot on an 8 GB M2 MacBook Air, offline.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Screenshot + questions in, calibrated typed decisions out. Locally, on Apple silicon. 0.9 s per decision on a 1080p screenshot on an 8 GB M2 MacBook Air, 2.1 GB of memory, $0 per call. The same screenshot costs $5.38 per 1,000 decisions in image tokens alone on Claude Sonnet 5.

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

Tagscommunity apple-silicon local vision mlx

Works forJev for Classification Jev for Decision gating

Used inJev for Software & DevTools

Languages: Python
Platforms: macOS
Capabilities: Open sourceSelf-hostedRuns locally

Highlights

0.9 s per decision on a 1080p screenshot downscaled to 1024 px on an 8 GB M2 MacBook Air, at 2.1 GB peak Metal memory
Answers matched the lab's PyTorch reference on 67 of 67 fields; the maximum calibrated probability difference was 0.023
One forward pass with no generation or sampling; temperature scaling T=2.7478 is the lab's and never changes which answer wins
Serves POST /v1/systemone with 1-16 questions, 0-8 images as base64 data URIs, an 8,192-token request limit and up to 62 choice options
The lab's card reports ECE 0.066 versus Jev's 0.095, but WildJailBreak 64.48 versus 96.29; the README warns not to use it as a safety gate

Watch out

Apache-2.0. Apple silicon only, 0.8B judgment is weak and the README says not to use it as a safety gate; its confidence is not the same statistic as Jev's, so thresholds cannot be reused.

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.

Hacker News: A one-token logprob wrapper that also classifies webcam frames.

allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker News

More Jev resources like this

jevos

1.3k

A local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.

Jev Models & Reimplementations#community#python#open-models

Community

feder-cr/jevC++

ruling

A local Jev-compatible server that reads Choice, Score and Noul answers from the logits of any MLX checkpoint or OpenAI-compatible endpoint with no training, serving POST /v1/systemone as a drop-in for the official TypeSafe SDK.

Jev Models & Reimplementations#community#mlx#calibration

Community

bradAGI/rulingPython

verdict

A Python server in front of llama-server that reads one-token label probabilities and exposes them as POST /v1/systemone.

Jev Models & Reimplementations#community#python#open-models

Community

khimaros/verdictPython

AutoTrust JEV-27B

An open distillation of Jev 1.13 onto Qwen3.8-27B: a 108.9M-parameter decision block answers typed questions in one prefill pass and adds a POST /v1/decide route beside a full System 2.

Jev Models & Reimplementations#community#open-models#local

Community

AutoTrust JEV-27B-VL

The vision sibling: the same text-trained System 1 head on Qwen3.8-27B's image encoder, taking images zero-shot, ranked first of 20 vision decision models on the community Jev Decision Index 0.3.

Jev Models & Reimplementations#community#open-models#local

Community

AutoTrust JEV-9B

The 9B member of AutoTrust's Jev students: a 40.2M-parameter decision block on Qwen3.5-9B with 0.019 mean KL to Jev 1.13 and 70.7% HumanEval left untouched, small enough for one modest GPU.

Jev Models & Reimplementations#community#open-models#local

Community

From the Jev guides

Original write-ups on what this entry is for.

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 Decision-Model Price War

Oct 11, 2026

One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.

Jev Models & Reimplementations#decision-model#benchmarks#pricing

Read article

What Is Actually on Hugging Face

Oct 8, 2026

Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.

Jev Models & Reimplementations#open-models#hugging-face#local

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai