wfzyx/vonArchived
User repository on GitHub
The open-source System One decision model. Sub-15ms, non-autoregressive, local drop-in alternative to TypeSafe Jev.
- Forks
- 71
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
Von
An open, non-autoregressive System One model with published weights and a TypeSafe-compatible /v1/systemone API; the 395M-parameter OptionMarker checkpoint scores discrete and ordinal questions locally.
- Stars
- 917
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- wfzyx
- Last verified
- Sep 2026
About this project
An open-source, non-autoregressive System One decision model. Calibrated Choice / Noul / Score inference from a 395M ModernBERT encoder, one forward pass, CPU-served. Von playing Doom. Every action is one forward pass scoring six option descriptions against a text rendering of the depth buffer. Same shipped weights that answer routing questions; no policy network, no RL.
Quoted from the project's README. Read the full README.
Used inJev for Gaming
Highlights
- The 395M-parameter OptionMarker checkpoint runs locally on CUDA, ROCm, Apple MPS, or CPU and answers in sub-25 ms.
- The README headlines 91.23% accuracy on adversarial multi-hop benchmarks, while its own 49-task table reports 72.0% macro and 72.4% micro.
- Author-reported ViZDoom results give 9.00 average kills versus 5.62 for hosted Jev 1.13, at sub-18 ms local latency.
- Calibrated with joint cross-entropy and Brier-score loss at temperature 1.0367; training used 250,000 class-balanced examples.
- Installs as pip install von-sdk or npm install von-sdk and ships a Doom benchmark script for reproduction.
Quickstart
bash
pip install von-sdk
pip install git+https://github.com/wfzyx/von.gitWatch out
Apache-2.0-licensed. Benchmarks are author-reported with no matching result artifact in the repo; the model runs locally with about 1.5 GB of weights and is not affiliated with TypeSafe.
Questions about this entry
Short answers drawn from the linked project or page.
That is the design goal. von serve exposes a /v1/systemone endpoint byte-compatible with the TypeSafe specification, and both the Python and TypeScript SDKs take a base URL override, so the swap is configuration. The 395M ModernBERT encoder answers Choice, Noul and Score in one forward pass without generating tokens.
Be careful with them. The README headlines 91.23% on adversarial multi-hop benchmarks while its own 49-task table reports 72.0% macro and 72.4% micro, and the repo ships no matching result artifact for either figure. Everything is author-reported, so reproduce on your own cases before repeating it.
Sometimes, and the repo draws the line explicitly. At the measured 0.80 default, Choice keeps about 25% of items at 92.4% accuracy on 702 held-out items while the rest sit near 59%, and 0.82 is the lowest cutoff clearing 90% kept-accuracy at 90% confidence. Out of domain the ranking itself breaks and no cutoff rescues it.
Long states are middle-truncated at 8,192 tokens by default so the question and options always fit, and --on-overflow refuse returns HTTP 422 instead, which is the no-truncation rule the project follows. Language is the harder wall: it is English only, and other languages get token matching with confidence it has not earned.
pip install von-sdk or uv add von-sdk for Python, bun add von-sdk for TypeScript, or docker run ghcr.io/wfzyx/von:cpu. The roughly 3 GB of weights are fetched from the Hub on first use into HF_HOME, and von serve auto-selects CUDA, MPS, OpenVINO or plain CPU.
It is deterministic multi-step computation for temporal or numeric states, driven by Von's own Choice decisions and zero generated tokens. It fires on computable structure such as two dates, a date and a duration, or two amounts, never on the question's wording, and runs fixed operators like add_duration, elapsed_hours and prorate.
Run von calibrate labels.jsonl, which refits the confidence map on frozen weights on CPU in minutes, writes marker_calibration.json into the checkpoint directory, and picks a scalar or feature map by k-fold CV while reporting NLL and ECE. Temperature never changes an answer, only how sure Von claims to be.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Von is posted as a sub-15 ms local drop-in for hosted Jev.
0x19975 pointsSub-15ms, non-autoregressive, local drop-in alternative to TypeSafe JevRead the thread on Hacker NewsX: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsMore Jev resources like this
An open 0.6B replica of Jev that turns states and questions into full probability distributions without decoding answer tokens, trained and evaluated on Maze, Snake, and ViZDoom.
Jev Repos & SDKs#community#python#open-models
Community
TianyuCodings/NanoJevPython
TianyuCodings/NanoJevPythonAn open 0.6B System One decision model that reads unstructured state and typed questions in one forward pass, returning calibrated distributions with zero decoded tokens; weights are published on Hugging Face.
Jev Repos & SDKs#community#python#open-models
Community
malevrigns/agent-jevPython
malevrigns/agent-jevPythonA reproduction of Jev that turns any Qwen checkpoint into a decision model serving the same /v1/systemone schema (Choice, Score, Noul), with no training and no generated answer text.
Jev Repos & SDKs#community#python#open-models
Community
zhengxuyu/litjevPython
zhengxuyu/litjevPythonOpen multilingual System 1 decision models with published checkpoints for choice, score and noul questions, plus a router that dispatches each request to the right checkpoint in one forward pass.
Jev Repos & SDKs#community#python#open-models
Community
NandhaKishorM/layaPython
NandhaKishorM/layaPythonA local server and Python package that runs open models with Hugging Face Transformers or PyTorch and returns Choice, Score, and Noul answers read from next-token logits instead of generated JSON.
Jev Repos & SDKs#community#python#open-models
Community
featherless-ai/simple-jevPython
featherless-ai/simple-jevPythonTurns local text and vision models into Jev-style decision models by reading probabilities from prefill logits alone, with SGLang, Transformers, and Apple Silicon MLX backends.
Jev Repos & SDKs#community#python#open-models
Community
Yinsongxu/LLM2JevPython
Yinsongxu/LLM2JevPythonFrom 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.