User repository on GitHub
Image inputs for Laya: calibrated, non-generative typed decisions over images + text (SmolVLM-256M backbone)
- Forks
- 8
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Models & Reimplementations
Laya Vision
An independent fork of Laya that replaces its text encoder with a small vision-language model, giving calibrated typed decisions over an image plus text in one forward pass from a 201M-parameter SmolVLM-256M backbone.
- Stars
- 91
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- r33drichards
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- The recommended checkpoint is thaitea/laya-vision, 201M parameters: A-OKVQA 59.8%, ScienceQA 82.4%, VQAv2 yes/no 71.4%, with a 41 ms median predict on an L4.
- Answers are choice, noul, and score over an image plus optional text; each option is cut to 48 tokens and the instructions get what is left.
- Inputs are cut to fit token budgets, and an answer whose question was cut carries a truncated field saying what was dropped; strict=True raises instead.
- It plays games from pixels: ViZDoom basic 0.99, Atari Freeway 0.81, Acrobot 0.76 on the autoresearch games benchmark where 0 is random play.
- Install is pip install -e . plus torchvision; ModernVBERT needs transformers 5.3 or newer; a free CPU Space answers in about 3 seconds per image.
Quickstart
python
import laya
from PIL import Image
agent = laya.load_vlm("thaitea/laya-vision")
result = agent.predict(
{"image": Image.open("photo.jpg"), "note": "customer says it arrived broken"},
{"category": {"type": "choice", "instructions": "What kind of item is this?",
"criteria": ["electronics", "clothing", "furniture", "food", "other"]}},
)
print(result["answers"]["category"])Watch out
README reviewed 2026-10-15. The repo calls itself an experimental research project, not affiliated with Convai Innovations, the authors of Laya. VQAv2 yes/no numbers come from an image-based re-split and are not comparable to published VQAv2 results, score answers are meaningful only on the two checkpoints trained on rubric data, and L4 latency varies about 50% between hosts.
Questions about this entry
Short answers drawn from the linked project or page.
An image-first decision model: you give it a picture, optional context, and questions, and it answers each as a multiple choice, a yes/no probability, or a rubric score in one forward pass with no text generation. It is Apache-2.0 licensed.
It is an independent fork of NandhaKishorM/laya that swaps the ModernBERT text encoder for a small vision-language model while keeping Laya's predict(state, questions) API, output schema, RLCD training objective, and temperature calibration.
The recommended thaitea/laya-vision, same weights as laya-vision-201m: 59.8% A-OKVQA, 82.4% ScienceQA, 71.4% VQAv2 yes/no, 41 ms on an L4. It is 15% smaller than the previous checkpoint and better calibrated, ECE 0.041 against 0.064.
It plays simple games from pixels, scoring 0.99 on ViZDoom basic and 0.81 on Atari Freeway on the autoresearch games benchmark, which reads 0.35 overall for the recommended checkpoint where 0 is random play and 1 is expert.
CPU inference works: a free Hugging Face Space answers in about 3 seconds per image on CPU, and a 41 ms median on an L4 GPU is documented. Install needs torchvision, and ModernVBERT requires transformers 5.3 or newer.
The recommended checkpoint answers 69.1% of 59,427 questions against 69.3% for the previous one. VQAv2 yes/no is a re-split and not comparable with published VQAv2 scores, and score answers only mean something on rubric-trained checkpoints.
Per-question predictions are committed under results/raw, and python benchmarks/verify_published.py recomputes them and checks them against the README table and each checkpoint's metrics JSON; the rules for changing numbers are in AGENTS.md.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: First among models under 10B on the Decision Index v0.2.1. Liquid AI's d1-3B is built from LFM2.5-VL-3B and…
X: The classic hotdog classifier as a real-time vision decision. Fayaz Ahmed rebuilt the not hotdog app on…
More Jev resources like this
A fork of Laya that answers choice, yes/no and score questions about an image plus optional text in one forward pass. This 201M checkpoint is SmolVLM-256M cut to 20 of 30 language layers and trained for two hours, with no text generation.
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
A 144.3M-parameter bilingual typed decision model fine-tuned from jhu-clsp/mmBERT-small: it answers noul, choice, and score questions in one forward pass with calibrated confidence and zero output tokens, and it runs on CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
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-jevPythonA Japanese System One decision model built on modernbert-ja-310m: send text and typed questions and get typed answers with probabilities from one forward pass, no text generation, at 314.6M parameters that also run on a CPU.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
hiroki-abe-58/sokudanPython
hiroki-abe-58/sokudanPythonOpen, CPU-sized Jev-style decision model: give it a situation, a question and options, and Decima returns calibrated probabilities in roughly 20-50 ms per decision as int8 ONNX, with no GPU, no API key and nothing leaving the machine.
Jev Models & Reimplementations#community#catalog-import#cobanov
Community
amyrmahdy/decimaPython
amyrmahdy/decimaPythonGestalt Labs' Jeff is a LoRA adapter on Qwen3-4B-Instruct-2507 that answers Jev-style choice, noul and score questions with labels and probabilities instead of prose, runs locally, and ships a train-your-own guide; on 9,730 fact-checking examples it trails Jev 1.13.0 on accuracy (0.8183 vs 0.8283).
Jev Models & Reimplementations#community#catalog-import#madewithjev
Community
From 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.