Organization repository on GitHub
Fast multi-modal decision model
- Forks
- 29
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
Intern-Decision
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.
- Stars
- 715
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- InternLM
- Last verified
- Oct 2026
About this project
🎮 Demo · 🤗 Model Weights. 简体中文 · Data interface · Reproduce evaluations. Intern-Decision turns a state, optional images, and multiple typed questions into decisions with probabilities. It fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector. Supported question types are choice, score, and noul (yes/no).
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Fine-tunes the Qwen3.5 language backbone while freezing the vision tower and projector, and choice, score, and noul answers share one causal forward pass.
- Ships training, native Hugging Face and XTuner inference backends, benchmark scoring, temperature presets, and a browser demo, but not the weights themselves.
- The 4B checkpoint reports 90.02% average accuracy across seven suites and 44.16 ms mean request latency on a single RTX 4090.
- A 96-case known-distribution calibration benchmark, its exact references, a deterministic generator, and a standard-library offline scorer ship alongside.
Watch out
Weights, training data, private calibration and validation records, images, and the preparation pipelines are not in the repo, so the released code alone will not reproduce training. The bundled calibration set is a small evaluation-only pilot with paired option variants rather than 96 independent trials. 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.
The README measures both in one table: Jev at 109.70 ms mean against 44.16 ms for the 4B checkpoint on an RTX 4090, with 289 input tokens and three fields answered in a single pass. Those are local BF16 numbers under one Transformers version, not an API comparison, and the weights are not in the repo.
Partly. The seven accuracy suites and the 96-case calibration benchmark ship with their exact references, a deterministic generator, and an offline scorer, and the scoring script needs only Python 3.10+ and the standard library. Training data, private calibration and validation records, images, and preparation pipelines are excluded.
Each checkpoint ships a temperature that was fitted on a separate calibration split, and the bundled benchmark was explicitly not used for fitting or selection; applying it preserved every predicted argmax label. If your data is off-distribution you would still want your own split, and inference only restricts logits to the legal answer symbols.
Linux, Python 3.12, and CUDA, with commands run from the repository root and Hugging Face and XTuner kept in separate virtual environments because their tested Transformers versions differ. The native backend loads Qwen3.5 directly and needs no XTuner import or compatibility patch.
Choice, score and noul, answered together in one causal forward pass. It fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector, so a state can include optional images alongside text.
No. The release covers training, two inference backends, benchmark scoring, temperature calibration and a browser demo, but model weights, training data, images and preparation pipelines are not included, so the released code alone will not reproduce training.
A 96-case pilot with six categories and 48 paired settings, where questions ask for random outcomes and exact reference distributions are stored separately from model inputs. It is evaluation-only: the README says do not fit temperatures or train on its references.
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/deepopenPython402GitHub stars
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.
Jev Models & Reimplementations#community#jevbest
Community
Zefan-Cai/Open-JevPython
Zefan-Cai/Open-JevPython363GitHub 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.