# Von

> An open-source, non-autoregressive System One decision model delivering calibrated discrete, probabilistic, and ordinal inference in sub-25ms, running locally without autoregressive text generation.

Von is an open-source, non-autoregressive decision model built on ModernBERT-Large (395M parameters) that performs classification, intent routing, and guardrail validation in a single forward pass — without the token-by-token generation overhead of large language models. Licensed under Apache 2.0, it runs entirely in-process or via an HTTP server and targets latency-sensitive pipelines where autoregressive LLMs introduce 500–2,000ms delays and nondeterministic schema parsing errors.

## What It Is

Von implements what its README calls the **System One** computational paradigm: reflexive, parallel, deterministic, and statistically calibrated decision-making. Rather than generating text, Von evaluates arbitrary discrete and continuous criteria directly over input state in a single forward pass. It exposes three core decision primitives — **Choice** (categorical classification), **Noul** (binary probability verification), and **Score** (ordinal continuous rating) — and supports evaluating multiple heterogeneous questions simultaneously via a `system_one` fan-out API. The model is named in recognition of John von Neumann and Ludwig von Mises, two foundational figures in computation and decision theory.

## Architecture and Training

Von-1.0 is built on **ModernBERT-Large**, pretrained on 2 trillion tokens of general web text, technical literature, and code. It is fine-tuned on a ~290,000-example balanced multi-domain corpus spanning operational/enterprise workflow (~25%), security/DevOps/compliance (~20%), safety/policy/moderation (~15%), linguistic/content semantics (~15%), triage/services (~10%), and adversarial reasoning anchors (~15%).

Post-training uses **Reinforcement Learning with Calibration Distribution (RLCD)**, minimizing a composite loss of Cross-Entropy and Brier Score (λ=0.5) to simultaneously optimize classification accuracy and probabilistic calibration. Temperature scaling converges at T=1.1692, yielding near-ideal expected calibration error (ECE). The result is output probabilities that reflect true predictive confidence rather than overconfident point estimates.

## Benchmark Performance

Von's README reports the following benchmark results across two independent evaluation suites:

- **jabr v2 benchmark** (49 tasks, 869 cases, out-of-domain decision making): Von achieves 71.5% macro accuracy, leading all open local System One models. Choice routing reaches 83.4%, with symptom triage at 100.0%, home services at 95.7%, and city routing at 94.7%.
- **ViZDoom real-time robotics/gaming** (8-seed evaluation, zero-shot from structured scene text): Von achieves 9.38 average kills in Defend the Center at sub-18ms latency, outperforming TypeSafe's proprietary Jev 1.13 API (5.62 kills at ~115ms) by +66.9%.

The benchmark table in the README compares Von against GLiNER2, finetuned Qwen3.5 4B, Laya, and TypeSafe Jev across model size, accuracy, latency, and hosting model.

## Hardware and Deployment

Von runs hardware-agnostically with native kernel optimization across:
- NVIDIA CUDA
- AMD ROCm (Linux)
- Apple Silicon Metal Performance Shaders (MPS)
- Multithreaded CPU

It can be deployed in-process via the Python SDK (`pip install von-sdk`) or TypeScript/JavaScript SDK (`bun add von-sdk` / `npm install von-sdk`), or as a production HTTP server via `von serve` compatible with the `/v1/systemone` wire protocol. Pre-packaged presets cover support ticket triage, email security/routing, content moderation, and security event triage. Composable patterns include confidence gating, route dispatch, composite risk scoring, and two-stage routing for high-cardinality taxonomies.

## Current Status

The repository was created on September 18, 2026 and last updated September 21, 2026, with 235 stars and 22 forks at time of indexing. The model weights are published on Hugging Face at `wfzyx/von-1.0`. The project is actively maintained under Apache 2.0 and positioned as a local, free drop-in alternative to the proprietary TypeSafe Jev API for System One decision workloads.

## Features
- Non-autoregressive single forward pass inference
- Sub-25ms GPU latency
- Choice: categorical classification with calibrated probability distribution
- Noul: binary probability verification with dual positive/negative framing
- Score: ordinal continuous rating via expected value
- system_one fan-out: evaluate multiple heterogeneous questions in one pass
- RLCD post-training with joint Cross-Entropy and Brier Score loss
- Temperature scaling for calibrated uncertainty (T=1.1692)
- Hardware-agnostic: CUDA, ROCm, Apple MPS, CPU
- Production HTTP server via `von serve` with /v1/systemone protocol
- Pre-packaged presets: triage, email, moderation, security
- Composable patterns: confidence gating, route dispatch, composite scoring, two-stage routing
- Python and TypeScript/JavaScript SDKs
- Apache 2.0 open-source license
- Hugging Face model weights (wfzyx/von-1.0)

## Integrations
Hugging Face, Python, TypeScript, Node.js, Bun, NVIDIA CUDA, AMD ROCm, Apple Metal Performance Shaders, ViZDoom, ModernBERT-Large

## Platforms
LINUX, API, DEVELOPER_SDK, CLI

## Pricing
Open Source

## Version
1.0

## Links
- Website: https://github.com/wfzyx/von
- Documentation: https://github.com/wfzyx/von#readme
- Repository: https://github.com/wfzyx/von
- EveryDev.ai: https://www.everydev.ai/tools/von-decision-model