SimpleJev.
GitHub Hugging Face
Researchers from

Open infrastructure Β· System 1 thinking models

Think fast.
Build open.

SimpleJev builds open infrastructure for System 1 thinking models, and does research on top of it. Models that read the situation and choose in a single step, with probabilities you can act on.

Explore JevAny Star on GitHub
Decision examples
2.18M+
Released models
5
Supported base LLMs / VLMs
26
Model families
7

Not every decision needs a long chain of thought

Large models reason step by step, which is slow and costly. Many decisions in agents, robots and software are fast and intuitive instead: route this ticket, pick this tool, take this action. A System 1 model sees the state and the options, then answers in one forward pass.

  1. S1

    Fast, calibrated choices

    Return a choice with probabilities, so your application can act or escalate to human review.

  2. S2

    Work alongside LLMs

    Let the LLM plan and recover; let Jev pick among the actions it proposes.

Jev inside LLM agent loops
One decision, one passIllustrative

STATE

"I was charged twice for the same order."

QUESTION

Which team should handle this ticket?

Choice & probabilitiesSingle forward pass
billing0.96βœ“
shipping0.03
account0.01

Our first release

JevAny

Project page

Infrastructure for training and deploying your own Jev model from open-source LLMs and VLMs.

Built with researchers from UIUC, Meta AI, MBZUAI and other institutions. We used JevAny to train and release five models on more than 2.18 million decisions. We hope it helps researchers and engineers train and deploy open Jev models faster, and speeds up research in this area.

Visit JevAny Code Models Tech report
  1. 01

    Data

    Turn raw logs into labelled decisions: state, question, options.

  2. 02

    Train

    Adapt any of 26 supported LLM and VLM backbones with LoRA.

  3. 03

    Evaluate

    Measure accuracy and calibration (NLL, Brier, ECE) on JevBench and Transfer.

  4. 04

    Deploy

    Serve a checkpoint behind one API, from Python or HTTP.

Results

All benchmarks

JevAny-Qwen3.8-27B beats Jev 1.13.0 on both controlled evaluations. At matched scale, our 4B and 27B releases also beat Kev, the most popular open Jev variant.

vs Jev 1.13.0

Accuracy Β· higher is better

JevBench

  1. JevAny-Qwen3.8-27B90.04%
  2. Jev 1.13.086.58%

Transfer

  1. JevAny-Qwen3.8-27B86.04%
  2. Jev 1.13.085.37%

vs Kev at matched scale

JevBench accuracy Β· higher is better

27B

  1. JevAny-Qwen3.8-27B90.04%
  2. Kev-27B85.28%

4B

  1. JevAny-4B Direct-Token80.95%
  2. Kev-4B75.32%

JevBench covers 231 public development items; Transfer covers 1,046 clean, knowable decisions. The 4B row is JevAny-Qwen3.5-4B-Direct-Token. These are diagnostic comparisons; see the evaluation protocol for details.

Released models

Hugging Face collection

What we work on

Infrastructure

Open tooling for the full loop: data preparation, training, evaluation and deployment of System 1 models.

Data & evaluation

Large-scale decision data and benchmarks that measure both accuracy and calibration.

System 1 + System 2 agents

When fast choices should replace LLM steps in agent loops, and when to hand control back.

Multimodal & embodied decisions

Decisions from text, images and video, across robotics, mobility, browsing and software.

An open, cross-institution team

SimpleJev brings together researchers and engineers from universities and industry labs. Everything we release is open: code, models and evaluation.

  • University of Illinois Urbana-Champaign
  • Meta AI
  • Mohamed bin Zayed University of Artificial Intelligence
  • and more
Open an issue Contribute