User repository on GitHub
Reverse engineering Jev using open models
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
Rev
A reverse-engineering attempt at Jev: three Qwen-based decision models with a LoRA tune plus pointer head that the README reports at 87.7-92.0% accuracy on a 975-question public benchmark, against 85.0% for hosted Jev.
- Stars
- 8
- Language
- HTML
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- robbalian
- Last verified
- Oct 2026
- Capabilities
- Open source
Highlights
- Reports 87.7%, 88.7% and 92.0% accuracy for the 4B, 9B and 27B versions against hosted Jev's 85.0% on the 975-question eval set.
- Speed and cost are server-to-server on one B200 per model on Modal; the 4B and 9B also undercut Jev's $0.0366 per 1k.
- The pointer head picks one of the offered options directly, so nothing is generated; training data is 19,792 public decisions from 18 datasets.
- Training took 33, 63 and 150 minutes for the three models, about $30 total; the whole project cost $438 on Modal.
- A Snake demo has the 27B averaging 187 moves per game against hosted Jev's 113, across 10 seeded games with a 1,000-move cap.
Watch out
Copy reviewed against the repo README on 2026-10-15; the pasted README was truncated inside the Snake section. Accuracy, speed and prices are the author's own measurements on Modal, and the eval set reuses validation splits of the training datasets, so no independent reproduction is cited.
Questions about this entry
Short answers drawn from the linked project or page.
A set of three decision models built on Qwen3.5-4B, Qwen3.5-9B and Qwen3.8-27B: each is a LoRA tune plus a small pointer head that picks an offered option directly, so nothing is generated. The repository publishes the models, training data and training scripts.
On the author's 975-question benchmark Rev reports 87.7%, 88.7% and 92.0% accuracy against Jev's 85.0%, with faster server-to-server speed; the 4B and 9B also come in under Jev's cost per 1k. These are self-reported results, not an independent reproduction.
From 19,792 public decisions taken from the train splits of RACE, CosmosQA, MultiRC, Social IQa and ContractNLI, plus short classification and policy rows from 13 other public sets. Training took 33 minutes for the 9B, 63 for the 4B and 150 for the 27B, one B200 each.
975 held-out questions from the validation splits of the same five core datasets, published and frozen in the repository's eval_sets/ directory. Accuracy, speed and cost are measured server-to-server on Modal with the GPU fully busy.
JSON output on hosted models was accurate but serial — 20 questions took 5.10 s against Jev's 0.27 s at 15x the cost per decision. Single-token outputs hurt accuracy without improving latency, and a head-only approach peaked around 80% on dev but did worse on external tests.
Playing one multiple-choice question per turn, the 27B survived 187 moves per game on average, the 4B 172, the 9B 144 and hosted Jev 113, across 10 games with a 1,000-move cap and identical seeds for everyone.
The README studies Jev's behaviour rather than its internals: question and token scaling stay flat, as does latency across input lengths, and prompt caching on repeated questions reduced Jev's latency, so the author added a cache-buster to every request to keep the comparison fair.
More Jev resources like this
An open, local attempt to bring Jev-style typed decisions to images: a Qwen3-VL backbone returns numerical distributions for independent choice, score and noul questions without generating answer text.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
andrueandersoncs/visual-jevPython
andrueandersoncs/visual-jevPythonAn open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.
Jev Models & Reimplementations#community#calibration#qwen
Community
denis-pplx/autojevPython
denis-pplx/autojevPythonA local vision decision tool built on Qwen3.5-4B: one forward pass turns an image, a question, and a set of options into option probabilities, with a measured 169 ms median complete request on a single RTX A6000.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
rjmswd/qwenjevPython
rjmswd/qwenjevPythonA training recipe for JEV-style System 1 decision models on Qwen bases: LoRA plus a 24-slot decision head distilled from Jev v1.13's output distribution, serving calibrated per-option probabilities through a TypeSafe-compatible endpoint.
Jev Models & Reimplementations#community#github#training
Community
shibing624/train-jevPython
shibing624/train-jevPythonJev-style medical QA implementation on Qwen3.5-4B that selects an answer from candidates in one forward pass instead of generating text. It reports 69.42% accuracy versus 67.41% for standard generation and a 10.37x speedup across 153,889 questions from nine benchmark sets.
Jev Models & Reimplementations#community#medical#benchmark
Community
Ztrura/Jev-MedQAPython
Ztrura/Jev-MedQAPythonAn open-source, Jev-compatible System One model: Dex answers choice, score and yes/no questions over text in a single forward pass from an Ettin-1B encoder, and speaks the same /v1/systemone shape as Jev.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
canadaornaw/dexPython
canadaornaw/dexPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.