User repository on GitHub
Jag: Direct Box Prediction for Efficient Visual Grounding
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Models & Reimplementations
Jag
Jag, a Jev-inspired visual grounding model: a Qwen3.5-0.8B backbone plus a lightweight MLP predicts normalized cxcywh bounding boxes for an image and referring expression in a single forward pass, removing autoregressive coordinate generation from the base model.
- Stars
- 10
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Last verified
- Oct 2026
About this project
Direct Box Prediction for Efficient Visual Grounding. Project page · Model · Demo. Jag adapts a compact multimodal model for continuous bounding-box prediction. Inspired by Jev, it returns the geometric answer directly, retaining the image–language processing of Qwen3.5-0.8B while removing autoregressive coordinate generation.
Quoted from the project's README. Read the full README.
Works forJev for Extraction
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- An MLP reads the last valid input-token representation and predicts normalized cxcywh coordinates in one forward pass, supervised by L1 and GIoU losses.
- The README reports 65.97 ms single-request latency and 15.16 samples/s for Jag versus 120.01 ms and 8.33 samples/s for NExT-Chat.
- RefCOCO-family accuracy at IoU >= 0.5: 91.28 on RefCOCO, 85.80 on RefCOCO+, and 87.72 on RefCOCOg, leading the compared models on three of five splits.
- Code is Apache-2.0; training uses Docker, ModelScope ms-swift, and EvalScope, with optional SwanLab logging.
- Training freezes the vision encoder and adapts a regression head before joint training of the language backbone and visual merger.
Watch out
Code is Apache-2.0, but model weights and datasets keep their respective licenses; it is a Jev-inspired adaptation rather than a Jev endpoint, and training or evaluation needs Docker plus NVIDIA tooling.
Questions about this entry
Short answers drawn from the linked project or page.
The README says Jag is inspired by Jev: instead of generating coordinates autoregressively, it returns the geometric answer directly from a compact multimodal model, predicting a bounding box in one forward pass.
Training and evaluation run through Docker Compose with the NVIDIA Container Toolkit and a GPU; inference can also be tried on the Hugging Face Space and the published xyzzzh/Jag weights linked from the repository.
At IoU 0.5 and above, Jag scores 91.28 on RefCOCO, 85.80 on RefCOCO+ and 87.72 on RefCOCOg, against a Qwen3.5-0.8B base of 79.74, 70.10 and 77.96. It improves on the base across all five test splits and leads the compared models on three of them, with LocateAnything ahead on RefCOCO and RefCOCOg.
Because coordinates are never generated token by token. Measured single-request latency is 65.97 ms at 15.16 samples per second and 2.52 GiB of GPU memory, against 120.01 ms and 8.33 per second for NExT-Chat and 239.05 ms and 4.18 for LocateAnything, making it 1.82x faster than NExT-Chat and 9.35x faster than Hi-Token.
An MLP reads the last valid input-token representation and predicts normalized cxcywh coordinates in one pass, supervised by L1 and GIoU losses. Head adaptation comes first, then joint training of the language backbone, visual merger and regression head, while the vision encoder stays frozen. Training uses ModelScope ms-swift and evaluation uses EvalScope inside Docker, with optional SwanLab logging.
Code is Apache-2.0, but the model weights and the datasets keep their own licenses. It is a Jev-inspired adaptation rather than a Jev endpoint, so nothing here calls a hosted model, and training or evaluation needs Docker plus NVIDIA tooling.
Clone the repo and use Docker: build the image, download the xyzzzh/Jag weights, then run scripts/infer.sh with an image path, a referring expression and a weight dtype such as bf16. The result is the bounding box in original-image pixels, and an online Hugging Face Space demo is also available.
More Jev resources like this
130GitHub stars
An 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/autojevPython36GitHub stars
Independent training toolkit and open weights for System One decision models in the same category as Jev: 17M, 68M, and 400M models answering Choice, Noul, Score, and reranking questions, with Python and ONNX inference and a documented generalization gap to Jev 1.13 on the project's S1MB benchmark.
Jev Models & Reimplementations#community#decision-models#onnx
Community
hotchpotch/bekko-system-onePython
hotchpotch/bekko-system-onePython3GitHub stars
A training-free runtime that turns dense Qwen3 or Qwen3.5 checkpoints into a Jev-style Noul, Choice and Score service, sharing one state prefill across isolated questions and reading each answer from candidate-token logits.
Jev Models & Reimplementations#community#qwen#local-inference
Community
HopLee6/QwevPython
HopLee6/QwevPython6.9kGitHub stars
A native MLX runtime for Laya's open typed-decision weights: short decisions on Apple Silicon, with no generated tokens, PyTorch, or cloud API.
Jev Models & Reimplementations#community#python#open-models
Community
mizorewww/laya-mlxPython
mizorewww/laya-mlxPython4.8kGitHub stars
An open-model semantic-if engine positioned as a local Jev-style decision stack, separate from the original SemIf repo but sharing the same author.
Jev Models & Reimplementations#community#python#open-models
Community
TheoLeeCJ/SemIf-OpenJevPython
TheoLeeCJ/SemIf-OpenJevPython2.2kGitHub stars
Open recipe for a local Jev-style decision model: a Qwen3.5-9B LoRA that scores one allowed answer token per Choice, boolean or rubric Score question, with data pipeline, training config and eval harness. Reports 90.1% on its 324-example holdout against Jev's 93.2%.
Jev Models & Reimplementations#community#lora#qwen
Community
bespokelabsai/nimblePython
bespokelabsai/nimblePythonFrom 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.