Von Model: 395M Open System One Model, Setup & Benchmarks

Von is an open-source System One decision model with 395M parameters, built on ModernBERT-large by the developer wfzyx. It is non-autoregressive: it scores every option in a single forward pass instead of generating text. It also serves a Jev-compatible POST /v1/systemone API, so it can stand in for TypeSafe Jev locally.


Maker	wfzyx
Architecture	ModernBERT-large encoder with an option-scoring head
Parameters	395M
Context	8,192 tokens
Current version	Von 1.2
License	Apache 2.0
Languages	English
Question types	`choice`, `noul`, `score`, plus multi-question `system_one`
Runs on	CUDA, ROCm, Apple Silicon (MPS), Intel (OpenVINO), CPU

Sources: Von on GitHub and the Von model card. Figures below are the project's own measurements.

How Von works

Von packs the state and every candidate option into one sequence. Each option gets a marker token, and the model reads a probability for each option off those markers. It never writes an answer out token by token.

Version 1.2 changed how options attend to each other. Each option now sees only the state and itself, never the other options, and position numbering restarts for every option. Reordering the options therefore no longer changes the answer. The project reports 0% answer flips under reordering, down from 49.5% in version 1.1.

Install and run Von

The SDK is on PyPI and npm:

pip install "von-sdk>=1.2.0"
# TypeScript / JavaScript
npm install von-sdk

Ask one question from Python:

import von

answer = von.decide(
    state="Order #123 was never delivered and the customer is asking for their money back.",
    choices={
        "refund": "Issue a refund",
        "track_order": "Help track the package",
        "escalate": "Escalate to a human agent",
    },
)

von.judge answers yes/no (noul) questions, von.rate answers ordinal (score) questions, and von.system_one answers several questions at once.

To replace a hosted Jev endpoint, run the Jev-compatible server and point your client at it:

von serve --host 0.0.0.0 --port 8000

A community MLX port also runs Von natively on Apple Silicon.

Von benchmarks

On the public JevBench tiers, Von 1.2 reports 100% on easy, 63.9% on standard and 38.7% on hard questions. The model card lists 36.9% for the hard tier, so treat that number as approximate. Von 1.1 reported 72.0% macro accuracy across 49 tasks (869 cases) on the project's v2 benchmark. It also reported 83.0% on choice routing.

For speed, the project advertises decisions in under 15 ms. Version 1.1 noted about 18 ms on a GPU. Measure on your own hardware, because latency depends on input length and the number of options.

Von vs Laya

Von and Laya are the two encoder-based System One models. Both are small enough to run without a large GPU:

	Von	Laya
Encoder	ModernBERT-large	ModernBERT-large (English), mmBERT (multilingual)
Parameters	395M	421M (English), 322M (multilingual)
Context	8,192 tokens	512 (English) or 1,024 (multilingual) tokens by default
Languages	English	English, plus 100+ languages
Jev API	`von serve`	laya-serve
Fine-tuning	Not the project's focus	Documented; see Fine-tune Laya

If your input is long English text, Von's 8,192-token context is its clearest advantage. If you need other languages or a documented fine-tuning path, start with Laya. Benchmarks from either project do not transfer automatically to your data, so compare both on your own labelled examples.

When not to use Von

Non-English input. The model card lists English only.
Bare labels. Von does best when each option has a descriptive criterion, not just a one-word label.
Open-ended answers. Von only scores options you provide. It cannot write text or reason step by step.

Von Model: 395M Open System One Model, Setup & Benchmarks

How Von works

Install and run Von

Von benchmarks

Von vs Laya

When not to use Von

FAQ

Is Von compatible with the Jev API?

Does Von run on CPU?

What changed in Von 1.2?