Vev

Vev: Jev-compatible, self-hosted decision model that answers yes/no, multiple-choice, or graded questions about text, JSON data, and image content.

Vev provides Jev-like decision models that accept text, JSON records, or images as input to answer specific questions. Users receive a probability score for every defined answer option in approximately 40 milliseconds, enabling automated judgment. The system includes two model sizes, vev-4b and vev-9b, both fine-tuned from Qwen3.5 for tasks like checking UI state or enforcing content policies. It serves the same API format as TypeSafe's Jev, allowing developers to switch models by updating the base URL in their existing SDK code. These open-weight models require an NVIDIA GPU, Python, and a CUDA build of PyTorch to run locally on private hardware.

Vev at a glance

Best for: Software engineer, Researcher
Runs on: API, CLI, self-hosted, open source
Pricing: Free
Released: October 2026
Category: AI research tool
Website: github.com/xiaooolong/vev

Key features

Supports yes/no, multiple-choice, and graded question formats
Processes images and text together in a single request
Returns probability scores for each defined option
Compatible with the Jev /v1/systemone API
Operates locally using open model weights
Supports rapid inference times of around 40 ms
Available in two sizes, 4b and 9b, fine-tuned from Qwen3.5

Use it for

Monitoring UI state in application screenshots
Checking images against safety policies
Verifying generated images against prompts
Identifying glitches in video game screenshots

Alternatives to Vev

NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
eu/jev — Inference hosted in Germany on Hetzner servers with no storage of prompts.
Kev — Provides calibrated probabilities and matches the TypeSafe API.
JEV sees — Turns visual input into program-ready structured choices and probabilities.
JevStation — Get typed decisions with probabilities instead of prose you have to parse.
Jev State — Test conversational decisions and export runnable TypeScript tests
Jiwo — It provides low-latency decision-making with a 50ms median request time.
SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
Decido — Runs locally on CPU while beating published benchmark scores.
Vegaml — Runs locally on your hardware with calibrated probabilities and conformal answer sets.
VLMEvalKit — Evaluate large vision-language models across numerous benchmarks with a single command.
Homi — It answers complex audio questions natively with high accuracy.

All Vev alternatives