Von
Apache-2.0 non-autoregressive System One model: one forward pass, order-invariant option scores, sub-25 ms local claims.
Overview
Von (github.com/wfzyx/von, Apache-2.0) is a compact non-autoregressive System One-style model: one forward pass scores premise text against explicit option descriptions for Choice, Noul, and Score-style tasks. README positions Von 1.2 as fixing option-order sensitivity from Von 1.1 (JevBench hard-tier shuffle diagnostic). Weights ship at huggingface.co/wfzyx/von with Python 3.12+ and TypeScript clients. Marketing copy cites sub-25 ms inference and Doom gameplay where movement picks are zero-shot from depth-buffer text, with tables comparing kills and latency to TypeSafe Jev 1.13 API and other open baselines on the author's hardware.
Problem: Autoregressive chat classifiers burn hundreds of milliseconds and KV cache RAM for routing tasks that only need scored options, not generated prose.
Built for: Builders who want Apache-2.0 weights, local sub-25 ms decisions, and order-invariant option scoring without TypeSafe API keys.
First indexed on Jev Directory: 2026-09-24
Creator and team
- Name
- wfzyx
How Jev is used
- Role in the product flow
- Bidirectional encoder scores each option against shared premise text in one pass
- Primitives
- ChoiceScoreNoul
- State in
- Premise string plus option descriptions; Von 1.2 isolates option tokens so scores do not depend on sibling option order.
- Decision out
- Discrete picks with calibrated-style probabilities per primitive; local in-process latency on GPU or CPU per README benchmarks.
- Load wfzyx/von weights from Hugging Face
- Format premise and option list per package examples
- Run single forward pass scoring all questions in parallel
- Threshold probabilities in your router or game loop
Von is the speed-first open cousin of hosted System One: no api.typesafe.ai round trip, no token-by-token decode for a letter answer. jaredpalmer-kev keeps Qwen autoregression but adds full fine-tune and SDK parity; togethercomputer-tev1 and featherless-simple-jev target different training and serving ergonomics. Laya and MLX ports covered in /learn/laya-vs-jev solve another local open stack; Von is explicitly tagged decision-model and System One in GitHub topics. Do not confuse with genai-craft-openvons. Pair with theoleecj-semif when you want a research server narrative, or classifier.dev when managed latency beats running 395M params yourself.
Sourced performance claims
- README cites Von 1.1 about 18 ms inference on documented JevBench and Doom tables versus TypeSafe Jev API about 115 ms in the same table footnotes.Source: github.com/wfzyx/von README benchmark tables
- Von 1.2 release notes claim option-order shuffle sensitivity dropped from 49.5% answer changes on hard tier to architecture-level invariance.Source: github.com/wfzyx/von README What's new in 1.2
- Public GitHub repo wfzyx/von had 635 stars and 46 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Apache-2.0 weights and Python plus TypeScript inference paths
- Order-invariant option scoring in Von 1.2 architecture
- Documented JevBench and ViZDoom-style zero-shot demos
- Sub-25 ms marketing claim for local interactive loops
- 250k example training story in README methodology section
Stack
- Python 3.12+
- TypeScript 5.x
- ModernBERT-family encoder
- Hugging Face Hub
Pricing: Open weights; inference cost is your hardware or cloud GPU time, not TypeSafe tokens.
Links
FAQ
- Is Von a drop-in TypeSafe API client?
- Von mirrors System One semantics in docs and examples but ships its own weights and clients. Point integrations at Von inference code, not api.typesafe.ai, unless you wrap it yourself.
- How is Von different from Kev?
- Kev is Jared Palmer's multi-size Qwen family with TypeSafe SDK compatibility and fine-tune tooling. Von is a smaller non-autoregressive encoder focused on fast local scoring.
- Should I trust the Doom kill counts as production proof?
- README presents them as zero-shot gameplay evidence on the author's GPU. Treat them as demos; run your own latency and accuracy harness before safety gates.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.