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

Lichen

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

Servicebeta
View on GitHub ↗

Overview

Lichen is a drop-in, API-compatible replacement for Jev that runs open-weight models on your own hardware. It serves the same /v1/systemone endpoint and returns the same choice, noul and score answers with a probability for every option, so existing Jev clients work unchanged. A request can also include images in the state, and questions are then answered about the pictures.

Using Google's Gemma 4 26B-A4B on a single 24 GB laptop GPU served from vLLM, the README reports 204 to 207 of the 231 public JevBench decisions correct, against 200 for Jev 1.13, at a median of 56 ms per decision. On images, it answered all 125 questions of a hand-checked Wikimedia Commons set and named the right road sign in 92 of 98 random photos, but when 'none of these' was offered it missed about a quarter of signs and was confident about most misses, which the README reports openly. The same vLLM server can host a 128k-token chat with thinking and images while it judges.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich team should handle this? / Which road sign is in the image?—
Yes / noDoes this convey urgency? / Is there a person in the image?—
ScoreAny ordered score question—

How it works

  1. Serve an open model with vLLM behind Lichen
  2. Send Jev-format requests with text or image states
  3. Lichen scores every option and returns probabilities
  4. Existing Jev clients consume the answers unchanged

Strengths

  • Runs locally on a single laptop GPU
  • Image support
  • Reported to match or exceed Jev 1.13 on public JevBench items
  • Honest reporting of confident errors

Caveats

  • Needs a 24 GB GPU
  • Overconfident when 'none of these' is offered on images
  • Public benchmark subset only

At a glance

Type
Service
Status
beta
Language
Python
Runtime
vLLM-backed HTTP server
Has code
yes
Stars
58

Use cases and integrations

How curated lists describe it

  • A local, API-compatible replacement for Jev, TypeSafe's System One model

    wh000wh000/awesome-jev-live

  • A local, API-compatible replacement for Jev, TypeSafe's System One model.

    hellogumbo/awesome-jev

  • Drop-in replacement serving `/v1/systemone` from open weights on your own hardware, returning the same choice, noul and score shapes.

    MrJev/awesome-jev

Frequently asked questions

Will my Jev client work with Lichen?

Yes, it serves the same /v1/systemone endpoint and answer shapes.

How accurate is it?

204 to 207 of 231 public JevBench decisions versus 200 for Jev 1.13, per the README.

Can it handle images?

Yes, images can be part of the state.

Related projects

  • 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
  • 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