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

Jev Local

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

Servicebeta
View on GitHub ↗

Overview

Jev Local (argos1111) is a local API that returns Choice, Score and Noul probabilities for questions about text, JSON and images using the same /v1/systemone shape as Jev. It offers three backends: LFM2.5 1.2B Instruct or LFM2.5-VL 1.6B reading the first answer token's log probability zero-shot (the default), Sarashina2.2 Vision 3B via llama.cpp, and a fine-tuned ModernBERT-Ja 310M cross-encoder that scores question-plus-state and candidate pairs.

The README reports JGLUE test results of 17 percent JNLI and 69 percent JComQA for LFM, 34 and 86 percent for Sarashina Q4, and 93 and 92 percent for ModernBERT trained on the train split, and cautions this is a practical comparison under one API rather than a fair ranking. It explicitly does not reproduce Jev's model, training or accuracy, and image input is a local extension. Setup scripts detect CUDA, ROCm or Metal, fetch and verify models, and fall back to CPU. Documentation is in Japanese.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich option fits this text, JSON or image?—
Yes / noIs this statement true of the input?—

How it works

  1. Run setup.sh to fetch a backend model
  2. Start the server with run.sh
  3. Send /v1/systemone requests
  4. Optionally add Sarashina or ModernBERT backends

Strengths

  • Fully local with automatic GPU detection
  • Image input support
  • Strong Japanese results with the fine-tuned backend

Caveats

  • Does not reproduce Jev's accuracy
  • Zero-shot LFM backend scores low on JNLI

Getting started

`git clone https://github.com/Argos1111/jev_local.git && cd jev_local && ./setup.sh --model text && ./run.sh --model text`.

At a glance

Type
Service
Status
beta
Language
Python
Runtime
Local server (llama.cpp, PyTorch)
Has code
yes

Use cases and integrations

How curated lists describe it

  • Replicating Jev with a local LLM.

    hellogumbo/awesome-jev

  • A local `/v1/systemone` server with two backends: an LFM model zero-shot, and a fine-tuned ModernBERT-Ja cross-encoder. Japanese documentation.

    MrJev/awesome-jev

Frequently asked questions

Is this Jev?

No, it is an unofficial adapter with Jev's API format.

Which backend is most accurate?

The fine-tuned ModernBERT-Ja on the reported Japanese tasks.

Does it support images?

Yes, as a local extension with the vision backends.

Related projects

  • Lichen★ 58

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

    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
  • AutoJev

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

    ModelresearchPython
  • 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