GitHub · Research and data
wfzyx/von
What the open alternative gives up and what it gains, in one table.
What it does
An open, non-autoregressive decision model — a 395M-parameter, 1.5GB local alternative — published with a head-to-head against the hosted model it substitutes: 72.0% macro accuracy on a 49-task suite against TypeSafe Jev's 96.6%, and 9.00 kills against 5.62 in the ViZDoom Defend-the-Center protocol: a clear loss on accuracy, a win on the real-time loop.
The repository's own description: “The open-source System One decision model. Sub-15ms, non-autoregressive, local drop-in alternative to TypeSafe Jev.”
How it works
A bidirectional ModernBERT backbone doing calibrated discrete, probabilistic and ordinal inference. The repository ships `training/`, `benchmarks/`, `tests/`, a Hugging Face build and a JavaScript bridge, and it names both evaluation suites it was measured on.
The repository, by the numbers
Read from the GitHub API on 2026-09-22. Stars and the last push move daily — quote them with the date, the way we do.
Numbers the author reports
These are the project's own figures, not our measurements — we keep ours on the statistics page and label them as ours. Read on 2026-09-23, with the sentence each figure appeared in:
- sub-25 msper inference — the open substitute's own claim, not Jev's
- “Calibrated discrete, probabilistic, and ordinal inference in sub-25ms.” — repository README
What we checked
- the tree: benchmarks, src/von, training, tests, hf, js, SOTA.md, CONTEXT_EXPANSION.md
- Apache-2.0 LICENSE and a pyproject.toml
- the comparison table in the README carries both models' accuracy and kill counts
- the GitHub description says sub-15ms while the README headline says sub-25ms and its table says ~18ms
What we did not check. We did not run the model or either suite; both sets of numbers are the authors'.