von
Apache-2.0 395M ModernBERT System One model serving calibrated Choice, Noul and Score locally as a drop-in /v1/systemone replacement.
Overview
Von is an open-source, non-autoregressive System One decision model: a 395M ModernBERT encoder that answers Choice, Noul and Score questions over JSON or text state in one forward pass, without generating tokens. It provides a drop-in server and client for the /v1/systemone protocol, ships as von-sdk for Python and TypeScript, and runs on CPU (OpenVINO), CUDA, ROCm and Apple MPS, including via Docker.
The lists describe sub-15 to sub-25 ms inference as reported by the author. Weights, training data, calibration and benchmarks are on the Hugging Face model card, and a guide covers fine-tuning on your own labels or another encoder. A demo shows the same weights playing Doom, scoring six option descriptions against a text rendering of the depth buffer each move, with no policy network or reinforcement learning.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Classify the root cause domain of this incident. | infrastructure · billing · feature_request |
| Yes / no | Is this blocking customers? | true · false |
| Choice | Which of six movement options to take in Doom? | — |
How it works
- pip install von-sdk or run the Docker image.
- Call von.decide or von.judge, or use the /v1/systemone server.
- Receive calibrated probabilities in one pass.
- Threshold and act; fine-tune on your own labels if needed.
Strengths
- Local, CPU-capable drop-in for Jev's API.
- Python and TypeScript SDKs.
- Open weights, data and calibration.
Caveats
- Weights are about 3 GB to download.
- Accuracy versus Jev depends on task; see the model card.
Getting started
pip install von-sdk (or docker run --rm -p 8000:8000 ghcr.io/wfzyx/von:cpu)At a glance
- Type
- Service
- Status
- beta
- Language
- Python, TypeScript
- Runtime
- CPU (OpenVINO), CUDA, ROCm, MPS server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Open alternative: a 395M non-autoregressive System One model that answers typed questions with calibrated probabilities in under 15 ms, positioned as a local drop-in replacement for Jev.
yibie/awesome-jevLocal non-autoregressive System One model with a `/v1/systemone`-compatible server and a Doom demo where each move is one forward pass
AnotiaWang/awesome-decision-modelsThe open-source System One decision model.
wh000wh000/awesome-jev-liveThe open-source System One decision model. Sub-15ms, non-autoregressive, local drop-in alternative to TypeSafe Jev.
hellogumbo/awesome-jevNon-autoregressive System One model with Python and TypeScript clients, published on Hugging Face under Apache 2.0. The author reports sub-25ms inference.
MrJev/awesome-jev
Frequently asked questions
Is it compatible with Jev clients?
Yes, it serves the /v1/systemone protocol.
Does it need a GPU?
No, it runs on CPU via OpenVINO as well as GPUs.
Can I fine-tune it?
Yes, docs/finetune.md covers training on your own labels.
Related projects
Open family of System One-style models fine-tuned from Qwen3.5 and Gemma that return typed decision probabilities in one pass.
ModelbetaPythonGemma 3 4B option scorer that prefills context once and scores all options in one pass, on MLX or PyTorch, with a Doom demo.
ServiceprototypePythonGemma 4 12B with its LM head replaced by a small answer head, serving Jev-style noul, choice and score decisions with custom code.
ModelresearchPythonIndependent MLX port of Laya for Apple Silicon: typed decisions in about 13 ms (7 ms multilingual) with no PyTorch or cloud API.
ServicebetaPythonDeploys DiffusionGemma-Jev behind a TypeSafe-compatible API on a Cloud Run GPU, with snake, dino and tetris demos.
ServiceprototypePythonMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython