User repository on GitHub
- Forks
- 56
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
Open-Jev
Decision models that return typed probabilities for candidate labels from a state, instructions, and options, shipped as LoRA adapters plus a trained decision head over pinned Qwen base weights.
- Stars
- 402
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- Zefan-Cai
- Last verified
- Oct 2026
About this project
Open-Jev-27B-v1.1 is here. Give your app a decision, with probabilities. Supply a context, questions and candidates; get typed probabilities directly, without autoregressive answer generation or parsing generated JSON. Watch the 32-second demo and try the interactive replay →. On X: new 27B v1.1 video · benchmark chart · poster and release links.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- The 2B, 9B, and 27B v1.1 packages carry LoRA adapters, a trained scalar decision head, and a saved calibration temperature; base Qwen weights are not included.
- The 27B v1.1 JevBench result is 197 of 231 including 80 of 111 on Hard, and the README states Jev stays ahead by three overall answers and one Hard answer.
- The opt-in Triton final-row kernel hit 2.63 to 2.75 times warm P50 on an H200 in isolation, but whole-model gains stayed near 1.0 without proven significance.
- A no-code workbench takes pasted text or a CSV and returns suggested labels with candidate probabilities, serving the 2B model on CPU.
Watch out
Packages ship adapters, a decision head, and a temperature rather than base weights, so you supply the pinned Qwen checkpoints and the loader yourself. The README is explicit about scope: the JevBench comparison covers only the 231 public tasks out of 534, fresh independent 27B results were still pending, and the acceleration opt-ins showed no established whole-model significance. No license is stated in the README section reviewed. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It is an independent implementation inspired by Jev that does not reproduce the proprietary training method, private weights, or training data, and claims none of the advertised speedups. On the 231-task public subset the 27B v1.1 scores 197 of 231 including 80 of 111 Hard, and the README states Jev remains ahead by three overall and one Hard answer.
Each package holds LoRA adapters, a trained decision head, and a saved calibration temperature, and the 27B v1.1 uses LoRA rank 8, so you need the pinned upstream Qwen weights plus the Open-Jev loader yourself. Consumer-hardware notes cover running the 9B in bf16 on a single 16 GB card.
Judge it from the whole-model numbers, not the operator ones. The isolated Triton kernel hit 2.63 to 2.75 times warm P50 on an H200, but released whole-model ratios were 1.004 to 1.020 with no established statistical significance, and an early 27B H100 smoke failed the unchanged-probability tolerance before full Torch and Triton-tail validation passed 336 comparisons.
The 27B v1.1 internal run covered 127,787 rows with zero failed, missing, or duplicate predictions under independent audit, and the external comparison covers the 231 public JevBench tasks out of 534 with no full-set score claimed. The videos are edited playback over synthetic held-out cases with differing candidate order, and the arcade replays recorded decision states rather than running a live rollout.
Yes, per the consumer-GPU notes: the 9B runs in bf16 on a single 16 GB card, with JevBench results reported on an RTX 4060 Ti, an RTX 5060 Ti and an H100, plus opt-in options that make the cached path faster. The hosted workbench is the opposite end, serving the released 2B on CPU, which the README says is slower than the GPU deployment measured.
On the v1.5.4 independent snapshot the 9B is #46 at 24.356 and the 2B #65 at 9.073, a chance-corrected competence rather than raw accuracy. In the natural-support pilot the released 2B scored 166 of 256, 64.8%, against BM25's 211 of 256, 82.4%, and the locked acceptance policy had 14 errors among 88 accepted rows, so the demo keeps human review as its default.
A fixed 64-step released-2B pilot improved selected synthetic Test from 65.6% to 77.3% and controlled OOD from 64.8% to 74.2%, but on the frozen 64-item public JevBench development sample Choice stayed 17 of 34 and Noul 9 of 23, and Score normalized MAE fell from 0.22393 to 0.20645 across only seven items. The README keeps those weaknesses on record.
More Jev resources like this
1.7kGitHub stars
Open-weight 4B and 9B models from TokenRhythm: NeoHorse-1 for tool use, coding, and instruction following, plus NeoHorse-Jev-4B for prefill-only Choice, Noul, and Score decisions.
Jev Models & Reimplementations#community#jevbest
Community
TokenRhythm/NeoHorsePython
TokenRhythm/NeoHorsePython1.5kGitHub stars
A 0.8B decision model that answers first and escalates only when unsure, plus nine LoRA adapters of roughly 41 MB each for guard, triage, tool choice, spam, and similar classification work.
Jev Models & Reimplementations#community#jevbest
Community
firelex/jeffPython
firelex/jeffPython2kGitHub stars
DeepOpen is a non-autoregressive decision engine built on Laya that classifies into developer-defined types in one forward pass, with a language router picking between English, multilingual, and typed-decision checkpoints.
Jev Models & Reimplementations#community#jevbest
Community
deepopen-com/deepopenPython
deepopen-com/deepopenPython715GitHub stars
Decision models that answer several choice, score, and noul questions over a state with probabilities in one forward pass, with training, two inference backends, calibration tooling, and the benchmark suite in the repo.
Jev Models & Reimplementations#community#jevbest
Community
InternLM/Intern-DecisionPython
InternLM/Intern-DecisionPython363GitHub stars
Open-weight 2B, 4B and 9B decision models that extend Jev's typed question format to photos, so an image plus a record can be checked against options you supply, with a probability on each and an explicit can't-tell. Runs on your own hardware.
Jev Models & Reimplementations#community#jevbest
Community
mohit67890/imajevPython
mohit67890/imajevPython253GitHub stars
A clone of browser-use's jev-ultrafast that swaps the hosted Jev decision call for Laya running locally through MLX, keeping the browser agent, action space and executor but making no paid decision API calls. Apple Silicon only.
Jev Models & Reimplementations#community#jevbest
Community
ipenywis/laya-ultrafastPython
ipenywis/laya-ultrafastPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.