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

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.

ModelbetaApache-2.0
Open source ↗

Overview

This is the Hugging Face model card for NeoHorse-Jev-4B, TokenRhythm's 4B structured decision model fine-tuned from NeoHorse-1-4B. Given a state and application-defined questions, it predicts decisions and probabilities for routing requests, selecting tools, checking conditions and rating outcomes, using prefill-only inference without autoregressive generation. Text requests can include multiple questions, and optional image requests combine a single image with text.

The card reports a 77.70 six-group text aggregate, highest among four open-weight decision models with complete results, and 83.26% mean accuracy on Nimble, VitaminC and MASSIVE. It also shows six decision demos: Tetris, Snake, robot manipulation, Mahjong, a four-player bomb arena and autonomous driving. Deployment options are vLLM, SGLang and a native Python, CLI and HTTP runtime serving /v1/systemone. The GitHub repository covers the wider NeoHorse family.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich candidate action fits this state?—
Yes / noIs this condition true?yes · no
ScoreExpected rating on defined levels—

How it works

  1. Download the weights from Hugging Face.
  2. Serve with vLLM, SGLang or the native runtime at /v1/systemone.
  3. Send text (or one image plus text) with questions.
  4. Use probabilities and expected ratings in routing rules.

Strengths

  • Open Apache-2.0 weights with image support.
  • Jev-compatible endpoint shape.
  • Demonstrated on games, robotics and driving tasks.

Caveats

  • Evaluations are by the model authors.
  • Image requests limited to a single image.

Getting started

Download TokenRhythm/NeoHorse-Jev-4B from Hugging Face and serve it with vLLM, SGLang or the native runtime.

At a glance

Type
Model
Status
beta
Language
Python
Runtime
vLLM, SGLang, HTTP server
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • TokenRhythm's Apache-2.0 4B decision model on NeoHorse-1-4B: text or a single image with text, prefill-only inference, served at `/v1/systemone`

    AnotiaWang/awesome-decision-models

Frequently asked questions

Can it take images?

Yes, optional requests can combine a single image with text.

Is it compatible with Jev clients?

It is served at /v1/systemone, the same endpoint shape.

What licence applies?

Apache 2.0.

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
  • 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