User repository on GitHub
对Qwen-3.5-4B这个视觉语言模型进行jev般改造，识别速度达到0.169秒/帧。切换其余Qwen模型直接可以按这个逻辑加速。
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
QwenJev
A 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.
- Stars
- 104
- Language
- Python
- Reading time
- 2 min
- Published by
- Community
- Author
- rjmswd
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceRuns locally
Highlights
- Measured on 2026-09-21: full-request median 168.694 ms, P95 173.989 ms, GPU model stage 137.378 ms, with 30 of 30 runs matching the expected answer.
- Conditions: one RTX A6000, BF16, batch=1, a 1280x720 image with one question and four options, 880 visual tokens and 989 total input tokens.
- One forward pass scores single-token answer letters and applies softmax; no token-by-token generation, no training, and no quantization of the BF16 weights.
- Fast kernels: causal-conv1d, the FLA gated delta rule, fused gated RMSNorm, and FlashAttention via PyTorch SDPA, with an operator audit in the results folder.
- Runs Qwen locally with no Jev API call and no TypeSafe key; the example classification is a fictional duplicate-charge support ticket that returns billing.
Quickstart
bash
python -m venv .venv
source .venv/bin/activate
pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu130
pip install packaging ninja
pip install --no-build-isolation -r requirements.txtWatch out
README reviewed 2026-10-15. The README itself calls the numbers a fixed-example latency measurement, not an accuracy benchmark, and notes probabilities are softmax scores without calibration. A CUDA GPU is required and the causal-conv1d build must match PyTorch, CUDA, and GPU architecture; no license is shown in the digest metadata.
Questions about this entry
Short answers drawn from the linked project or page.
A local visual decision tool built on Qwen3.5-4B: you give it an image, a question, and options, and one forward pass returns the options with probabilities. The madewithjev listing reports 0.169 seconds per frame.
The problem design draws on Jev's Choice pattern and public support-classification examples, but it runs Qwen locally: no Jev API call, no TypeSafe key, and no affiliation with TypeSafe. It is a Jev-style conversion of a vision-language model.
Median 168.694 ms and P95 173.989 ms for a complete request on one RTX A6000 with BF16 and batch=1, over 30 runs, all matching the expected answer. The README calls this a latency measurement, not a general accuracy evaluation.
An NVIDIA CUDA GPU, Python 3.11, PyTorch 2.11.0+cu130, Transformers 5.13.1, FLA 0.5.2, Triton 3.6.0, and a causal-conv1d build matching PyTorch, CUDA, and the GPU architecture; the program errors out rather than silently using slow kernels.
Use python benchmark_fast.py --model Qwen/Qwen3.5-4B --compile-language --output results/local-run, or add --local-files-only for a local copy. The output directory must not already exist, and the results land in summary.json and records.jsonl.
No. They are softmax-normalized scores over the single-token answer letters, and the README states no calibration was applied. Compilation can cause small numeric differences that are not guaranteed bit-identical, and nothing is reused across requests.
A fictional customer message with a duplicate-charge complaint is classified into billing, technical, sales, or other; the model returns billing. The image is fictional and the question and option definitions live in examples/support_ticket.json.
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
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…
X: Liquid d1 with vision
More Jev resources like this
A local, open-source personal AI assistant for your job: mail, chats and tickets become tasks, agents you already use such as Claude Code, Codex and Gemini do the work, and nothing is sent until you approve. The current release is v0.3.8.3.
Jev Tools & CLIs#community#catalog-import#logicrw
Community
ldbumble/taskuaryPython
ldbumble/taskuaryPythonA tuner that trains language models to answer closed-set decisions with a complete probability distribution in one forward pass, using a tokenized multi-class Brier loss so only the correct choice is needed to train.
Jev Repos & SDKs#community#catalog-import#madewithjev
Community
liushiliushi/jevtunerPython
liushiliushi/jevtunerPythonAn Ollama-style local runner for System One decision models from Hugging Face: pull, serve and query them behind the same /v1/systemone API as TypeSafe's Jev, with a terminal model manager and a browser playground.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
nvkudva/ollajevPython
nvkudva/ollajevPythonA local app that reads a few lines of notes about your product and lets Jev pick its way to a finished landing page: Jev never writes text, it only chooses between options, and code builds the page from those picks.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
sagapranav/jev-pagesHTML
sagapranav/jev-pagesHTMLAgent tooling that turns a model idea into a buildable LEGO model: the agent writes a plan and a generator script that produce LDraw source, plus a 3D viewer, a Blender-editable glTF file and a Meta Quest 3 VR view. Part search uses jev-rerank, a semantic search reranked by TypeSafe's Jev.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
anteloc/ldraw-novaPython
anteloc/ldraw-novaPythonAn enterprise admin-console test tool that generates replayable Excel test cases from live browser pages, or replays existing Excel cases, driving Playwright with local Laya or a compatible decision API to match fields, options, and targets.
Jev Tools & CLIs#community#catalog-import#madewithjev
Community
anthonyli/laya-pilotJavaScript
anthonyli/laya-pilotJavaScriptFrom 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.