RSI-Jev
Open Jev-style text-and-image decision models built by a self-improving AI research loop, with an effort setting that picks depth.
Overview
RSI-Jev is a recursively self-improving research loop, the next version of AutoScientists, that builds Jev-style System One models. AI agents propose hypotheses, register predictions before spending GPU time, run experiments and retire champions when evidence says so; every release and every failed experiment is published. The README reports 8 releases in 13 days and 496 experiments.
The resulting open models answer yes/no, pick-one or rubric questions about text, chats or images, returning a probability for every option in one forward pass with nothing generated. v6.1-VL has answer heads at several layers, and an effort setting chooses depth: low (layer 16, about 23 ms median), medium, high, or auto, which stops at the first confident layer. On the Decision Index 0.3 public set the README reports 65.64 for the 27B with effort auto and 50.98 for the 4B, with the official score pending. A server speaks Jev's /v1/systemone API so existing Jev clients work after changing the base URL, and an MLX path supports Macs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the user request a refund? | — |
| Yes / no | Does the photo show visible damage? | — |
| Choice | Pick-one-of-k questions about text or images | — |
How it works
- Serve a checkpoint with rsi-jev serve and an effort level
- POST state (text or images) and typed questions to /v1/systemone
- Model answers from the chosen layer depth
- Use probabilities in code; auto effort adapts depth per question
Strengths
- Open weights for text and images
- Effort/depth trade-off with low latency
- Jev-compatible API
- All experiments published, including failures
Caveats
- Official Decision Index score pending
- 27B needs a 96 GB GPU
- Self-reported results from an automated loop
Getting started
pip install "rsi-jev[fast,vision] @ git+https://github.com/Shanghua-Gao/RSI-Jev" then rsi-jev serve v6.1-vl-4b --effort auto --port 8000.At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- HTTP server (GPU or MLX)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev-like model built by an RSI system: an autonomous AI-scientist research loop that proposes, runs and judges every experiment and publishes all of them (496 experiments, 8 releases in 13 days) — whose open models answer `Noul`, `Choice` and `Score` about text and images on the same `POST /v1/systemone` wire format, returning a calibrated probability for every option from one forward pass; v6.1-VL 4B scores 50.98 and v6.1-VL 27B 65.64 on the Decision Index 0.3 public set, the 27B with effort `auto`, which picks its depth per question.
yibie/awesome-jevJev-like models trained by a recursively self-improving (RSI) AI research system that publishes every experiment, failures included. 4B and 27B (v6.1-VL), with a probability for every Choice, Score, and Noul option in one forward pass over text or images, an effort setting (low, medium, high, or auto, which picks its depth per question), open weights, and a `/v1/systemone`-compatible server.
AnotiaWang/awesome-decision-modelsTyped-decision models (noul / choice / score) trained by a self-improving loop…
wh000wh000/awesome-jev-liveOpen 4B and 27B models (v6.1-VL) that answer Choice, Score, and Noul about text and images on the same `POST /v1/systemone` wire format, returning a probability for every option in one forward pass at a chosen effort (low, medium, high, or auto, which picks its depth per question), trained by a self-improving loop of agents that publishes every experiment it ran, including the ones that failed.
cobanov/awesome-jevTyped-decision models (noul / choice / score) trained by a self-improving loop of AI agents — checkpoints, the code that produced them, and every version that failed.
hellogumbo/awesome-jevTyped-decision checkpoints produced by a self-improving loop of agents that register predictions before spending GPU time, published with the code that produced them.
MrJev/awesome-jev
Frequently asked questions
What is the effort setting?
It picks how many layers answer: low is fastest, high deepest, and auto stops at the first confident layer.
Can existing Jev clients use it?
Yes, the server speaks Jev's API; change the base URL.
Who built the models?
An automated AI-scientist loop that runs and publishes every experiment.
Related projects
Independent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPythonJev-style choice-plus-probabilities from GLM-5.3-Flash in one forward pass, in Edgeless Systems' confidential-computing API.
ServicebetaPythonSemIf (formerly OpenJev): reads typed option probabilities directly from open models like Qwen3.5-4B, including image inputs, on a home GPU.
ModelresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython