Can existing TypeSafe SDK code run against Vev unchanged?

If the code uses the /v1/systemone request format, yes: point the client's base URL at a local vev serve process on 127.0.0.1:8009 and pass any placeholder API key. The README demonstrates the official Python SDK used this way with Choice and Noul questions and image states.

Is Vev the same as TypeSafe's Jev?

No. Vev is an independent research preview that speaks Jev's wire format and is not affiliated with or endorsed by TypeSafe. Its probabilities rank answers well but are less calibrated than Jev's, with expected calibration error from 0.01 to 0.14 depending on task, and a chosen threshold should be validated on your own labelled data.

What hardware and runtime does Vev need to run locally?

Python 3.11 or newer, an NVIDIA GPU and a CUDA build of PyTorch; CPU and Apple Silicon are untested. Install with pip install git+https://github.com/Xiaooolong/vev, then vev serve downloads the weights on first start and listens on 127.0.0.1:8009.

How big are the Vev models, and can I run just the adapter?

vev-4b needs about 10 GB of GPU memory and vev-9b about 19 GB. Passing --model CountingSheep/vev-4b-lora downloads only the 130 MB adapter and applies it to Qwen/Qwen3.5-4B, which is reused from the Hugging Face cache, and --revision v0.1.0 pins the weights.

Are Vev's weights licensed for commercial use?

No. The repository code is Apache-2.0, but the published weights are non-commercial only (CC BY-NC 4.0). Vev is an independent research preview and is not affiliated with or endorsed by TypeSafe.

How does Vev handle many questions or concurrent requests?

All questions in one request run as a single batch, so on one H800 a request of 100 short-text questions finished in 232 ms. A vev serve process handles one request at a time and queues the rest, so for more throughput run one process per GPU behind a load balancer.

Can I use Vev without the HTTP server?

Yes. The CheckpointEngine runs offline, for example to score a large file: engine = CheckpointEngine("CountingSheep/vev-4b") then engine.run(state, questions) returns the probabilities. Another inference engine can follow the prompt template and token-probability notes in spec/systemone-api.md.

CommunityJev Models & Reimplementations

Vev

Jev-compatible decision models with vision input: LoRA fine-tunes of Qwen3.5 at 4B and 9B serving Choice, Score and Noul on the /v1/systemone wire format, with screenshots and photos placed beside text in the state. Runs on your own NVIDIA GPU; weights are non-commercial.

See highlights

Open github.com

Stars: 17
Language: Python
License: APACHE-2.0
Published by: Community
Last verified: Oct 2026

Xiaooolong/vev

User repository on GitHub

View on GitHub

Jev-like decision models that can also see images — open weights on Qwen3.5, run on your own GPU

Forks: 2
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Jev-like decision models that can also see images. vev-4b playing Doom in real time. On each look, a small harness cuts the screen into 8 vertical slices and asks Vev one yes/no question per slice in a single request ("does this slice show a monster?"); a fixed rule turns the answers into turning and firing. The game never waits for the model.

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

Tagscommunity vision open-weights local-inference

Works forJev for Classification Jev for Scoring Jev for Verification Jev for Moderation

Used inJev for Software & DevTools Jev for E-commerce & Retail Jev for Gaming Jev for Media & Entertainment

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

Highlights

vev-4b needs about 10 GB of GPU memory; vev-9b about 19 GB, both on NVIDIA GPUs.
Answers come back in roughly 40 ms and generate no text tokens.
Serves the same /v1/systemone format, so existing Jev SDK code works after a base-URL change.
UI-state questions scored 0.928 and 0.951 accuracy with vev-4b; policy checks 0.742.
On one H800, 100 short-text questions ran as one batch in 232 ms.

Watch out

The repository code is Apache-2.0 but the published weights are non-commercial only (CC BY-NC 4.0), and v0.1 is a research preview reported as less accurate than Jev on most text sets. Option order can flip the top answer on 13 to 17 percent of questions, and CPU or Apple Silicon is untested.

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

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

Qwev

A training-free runtime that turns dense Qwen3 or Qwen3.5 checkpoints into a Jev-style Noul, Choice and Score service, sharing one state prefill across isolated questions and reading each answer from candidate-token logits.

Jev Models & Reimplementations#community#qwen#local-inference

Community

HopLee6/QwevPython

AutoJev

130

An open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.

Jev Models & Reimplementations#community#calibration#qwen

Community

denis-pplx/autojevPython

RSI-Jev

109

Open decision models from a self-improvement loop that publishes every experiment, failures included. They answer Noul, Choice and Score questions on the same POST /v1/systemone wire format as Jev, returning a calibrated probability per option in one forward pass; the 27B scores 65.64 on the Decision Index 0.3 set.

Jev Models & Reimplementations#community#decision-models#system-one

Community

Shanghua-Gao/RSI-JevHTML

Gutsy

An Apache-2.0 0.8B decision model that returns calibrated probabilities for yes/no, choice and score questions on CPU. Distributed as GGUF for llama.cpp and served behind a Jev-style /v1/systemone API, so existing Jev clients mostly need only a new base URL.

Jev Models & Reimplementations#community#gguf#local-inference

Community

kouhxp/gutsyHTML

OpenDecision

An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.

Jev Models & Reimplementations#community#python#local

Community

deepanwadhwa/OpenDecisionPython

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