Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

AutoJev

Open 27B multimodal decision model built largely by autonomous agents, serving choice, noul and score on a TypeSafe-compatible API.

ModelresearchMIT (code), Apache-2.0 (weights)
View on GitHub ↗

Overview

AutoJev-27B is an independent, open-weight multimodal decision model inspired by Jev and based on Qwen3.8-27B. It returns probabilities over supplied choices in one forward pass per question and serves choice, noul and score on a TypeSafe-compatible /v1/systemone endpoint, with optional base64 images and a browser playground. The author says autonomous agents executed the research, data generation, training, evaluation and deployment while the human set goals and scope.

Training used full-weight SFT on one H200 with 73,000 unique examples over 286 updates; the released model is checkpoint 200, with temperature calibration fitted separately. The README's results table reports 84.60% accuracy (ECE 0.04282) against 82.79% for Jev and 69.83% for the base model on its own evaluation. The curated training corpus is not bundled. Running it needs a GPU with room for about 49 GiB of BF16 weights.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich of the supplied options is correct?—
Yes / noIs this statement true?yes · no
ScoreWhere on this ordered scale does the state fall?—

How it works

  1. Download weights from Hugging Face
  2. Start autojev-serve
  3. POST to /v1/systemone with state, questions and optional images
  4. Use calibrated probabilities in code

Strengths

  • Open weights (Apache-2.0) and MIT code
  • Image input supported
  • Drop-in for TypeSafe clients

Caveats

  • Needs about 49 GiB of GPU memory
  • Benchmark is the author's own; training data not released

Getting started

git clone https://github.com/denis-pplx/autojev.git && cd autojev && uv sync --frozen --python 3.12 && uv run hf download denis-pplx/autojev-27b --local-dir checkpoints/selected && AUTOJEV_CHECKPOINT=checkpoints/selected uv run autojev-serve

At a glance

Type
Model
Status
research
Language
Python
Runtime
local GPU server
Has code
yes
License
MIT (code), Apache-2.0 (weights)

Use cases and integrations

How curated lists describe it

  • Open recipe: 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 calibration plots released.

    yibie/awesome-jev

Frequently asked questions

How does it compare to Jev?

On the author's evaluation, 84.60% accuracy versus 82.79% for Jev.

What hardware do I need?

A GPU with space for about 49 GiB of BF16 weights plus overhead.

Does it need a TypeSafe key?

No, it is self-hosted.

Related projects

  • Lichen★ 58

    Drop-in local replacement for Jev that serves /v1/systemone from open weights, including questions about images.

    ServicebetaPython
  • Jev Local

    Local /v1/systemone server in Japanese docs with LFM, Sarashina and fine-tuned ModernBERT-Ja backends, including image input.

    ServicebetaPython
  • NeoHorse-Jev-4B

    Hugging Face card for NeoHorse-Jev-4B: Apache-2.0 4B decision model for text or image-plus-text, served at /v1/systemone.

    ModelbetaPython
  • Standard One

    Standard Thinking's Apache-2.0 3B and 8B decision models on Ministral 3 that score bounded answers via /v1/systemone, with vision.

    ServicebetaPython
  • Winnow

    Apache-2.0 Gemma 4 12B fine-tune for local typed decisions, chat and image input, served via /v1/systemone and chat completions.

    Serviceresearch
  • JEV-27B-VL-Uncensored-GGUF

    Static GGUF quantizations by mradermacher of SargeDev's JEV-27B-VL-Uncensored vision decision model.

    Model