↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Hybrid execution/

Von

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~986 words)
agent-curated 512 cpu-inference 1 decision-models 20 fully-ai-generated 615 hybrid-execution 25 llm=glm-5.3-flash 523 open-weights 15 research-note 482 structured-outputs 27 system-one-models 19
Table of Contents

Von is wfzyx’s Apache-2.0 non-autoregressive decision model, a 395M ModernBERT encoder that answers Choice, Noul, and Score questions over JSON or text states in one forward pass with no generated tokens, shipped as a TypeSafe-wire server plus Python and TypeScript SDKs and runnable on CPU through OpenVINO.

Von is the wave’s smallest credible generalist, and its third-party benching says exactly what its size predicts: the cheapest, fastest CPU deployment in the category, English-only in practice, and near the bottom of the accuracy tables the decoder fine-tunes lead.

What it is
#

A ModernBERT encoder (about 3 GB of weights on Hugging Face) that reads option descriptions against a shared state and scores them in one pass: Choice over K described options with a probability over all of them, Noul as a probability, and Score as a calibrated position on an ordinal scale. It ships as von-sdk for Python and TypeScript, a Docker image (ghcr.io/wfzyx/von:cpu), and runtimes for CPU (OpenVINO), CUDA, ROCm, and Apple MPS, speaking TypeSafe’s /v1/systemone wire protocol so SDK code written against the hosted contract runs unchanged. The README leads with a Doom demo whose every action is one forward pass scoring six option descriptions against a text-rendered depth buffer, on the same shipped weights that answer routing questions, and the model card carries the training data and the benchmark record.

Status
#

Active and community-scale: the repository was created 2026-09-18 and pushed 2026-10-08, with 880 stars as of 2026-10-11, and the checkpoint shows 103,556 downloads and 57 likes (GitHub and Hugging Face APIs, as of 2026-10-11); von-sdk 1.3.7 has been on PyPI since 2026-10-02.

Star History Chart

The scoreboard readings are humbling: von-1.3 scores 8.17 on the Decision Index 0.3.1 board against Jev’s 60.11, and Von does not carry a JevBench composite row. The third-party local-jev-bench suite (measured 2026-10-05) states the model card’s own boundary in its findings: English-only, 15 percent on a Korean set, near the bottom of its accuracy tables on every set but the small English one, at 32 to 411 ms first-call p50; Ollaya’s table scores von at 0.447 on typed decisions with 8,192-token states at 23 ms.

Strengths
#

  • CPU-first serving is the differentiator that matters: 20 to 30 ms decisions on commodity cores put this contract inside processes no GPU touches, and the Docker and dual-SDK packaging makes that a ten-minute install.
  • The 103k download count on a 395M checkpoint says the size niche is genuine, and von is the only encoder-scale generalist in the wave besides Laya’s family.
  • TypeSafe-wire compatibility plus Python and TypeScript SDKs keep it drop-in for code written against the hosted contract.
  • The model card publishes the training data and benchmark record rather than a claims table, which is more than most of the wave’s launch posts offer.

Cautions
#

  • Accuracy is the price of the size: near-bottom readings on every third-party table that includes it, and a Korean-language 15 percent that its own card’s English-only note predicts.
  • No JevBench composite row exists, so its calibration has no independent reading beyond local-jev-bench’s distribution-distance work.
  • One maintainer, 880 stars, and an SDK versioned in the ones digit: the bus factor and API-stability cautions that apply to every one-person entry in this wave apply here.
  • The 395M encoder’s ceiling is architectural; fine-tuning recipes exist in the repo, but no published fine-tune has moved it into the top group.

Pricing
#

Free and open: Apache-2.0 code and weights, no hosted service and no paid tier. The cost is about 3 GB of weights and whatever CPU you already have.

Compared to
#

  • Laya: the other encoder-scale family; Laya ships three checkpoints and a router with working multilingual coverage and its own fine-tuned typed-decisions specialist, Von ships one smaller model with a CPU-first runtime story.
  • Jev: the closed original; against hosted Jev, Von trades most of the accuracy for hardware the size of a laptop, which is the trade the contract’s millisecond pitch implies but no decoder entrant actually delivers.
  • Ollaya: the runtime that packages von and its calibration, and whose own table is the fastest way to see what the 395M ceiling costs.

Bottom line
#

Recommended for embedded, on-prem, or per-process decision layers where a GPU is off the table and English-only, mid-accuracy decisions are acceptable. Not for accuracy-led rankings, multilingual states, or anyone who needs a calibrated threshold validated by a second party. The disagreeable claim I will defend: Von’s near-bottom accuracy and six-figure downloads are the same fact, the demand in this category is for the deployment model first and the scoreboard second, and every accuracy-led entrant that ignored that has grown slower than the CPU ones.

Changes
#

  • 2026-10-11 - Created from the entrant scan (awesome-jev cross-check), accepted on the CPU-first niche (880 stars, 103k checkpoint downloads) with the near-bottom third-party accuracy readings recorded as the governing caution.
  • 2026-10-11 - Added the star history chart.

See also
#

  • Laya - the other encoder-scale family, multilingual and routed where Von is single and CPU-first
  • Jev - the closed contract Von reproduces on CPU-class hardware
  • Ollaya - the runtime that packages von and carries its typed-decisions row
  • Hybrid Execution Feature Matrix - the category comparison this note joins

References
#