Open source, self-hosted. There is no hosted API. You download the weights and run them on your own hardware. How open-source models are listed.
What Von is
Von is an open-weights decision model published by wfzyx, the GitHub account of a solo developer, Victor Hugo. The repo opened on 18 September 2026 and the weights went up on Hugging Face the next day.
Von is archived. The author added an archived banner on 17 October 2026 and says the weights will not be updated. The README and model card point new projects to LiquidAI/d1-omni-600M, one of the open versions on the d1 page. The weights, SDKs and benchmarks stay up for existing users.
How it works
Von is a fine-tuned ModernBERT-large encoder with a small scoring head, 395 million parameters in all. The state, the question and every option go into one sequence. Each option gets a marker, and the head scores the markers in one pass, so nothing is generated. It answers Choice, Noul and Score questions with a probability over the options. Since version 1.2 each option is scored on its own, so the order of the options does not change the answer.
von serve exposes /v1/systemone, which the README calls byte-compatible with TypeSafe’s spec. Code written for Jev can switch by changing the base URL. It runs on CPU through OpenVINO, and on CUDA, ROCm and Apple chips.
What it is good at
It is small and fast on cheap hardware. The Jev Decision Index measured a median of 17.7ms per request. The author reports 96ms on a 4-vCPU server and 72.0% on the 49-task jabr v2 suite; those numbers are author-run. The SDK includes a confidence gate that splits answers at 0.80 into automatic and escalate, and the README shows Von picking moves in Doom.
What it is not for
Its accuracy is low on broad tests. In edition 0.4.1 the Jev Decision Index gives it a Full score of 11.65 and 13.86 on the public part, with a calibration error of 0.2157: mean confidence was 0.635 against 0.419 accuracy. The author’s own JevBench v1.4 table puts Von 1.2 at a composite of 27.5 against 63.3 for Jev 1.13. JevBench now lists only a carried v1.5 score for it.
The model card lists the limits. It is English only. It is weak on long policies and multi-step date or number questions. Its calibration was fitted on JevBench public items and does not carry to other data, so refit it with von calibrate before acting on the confidence. And with no maintenance, known gaps will not be fixed.
Access
There is no hosted API or price. Install von-sdk from PyPI or npm (last release 1.3.7, 2 October 2026), or run the ghcr.io/wfzyx/von Docker image. The repo had 919 stars on 20 October 2026. For a maintained open model, the author points to d1-omni-600M; Laya is another ModernBERT-large option.
Specifications
| System One Score | 16.3 from 1 of 3 boardsJev Decision Index 16.3. How the score works |
| Question types | ChoiceNoulScore |
| Max Choice options | Not documented |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 8,192 tokens |
| State budget | 8,192 tokens |
| Rate limit | None. It runs on your own hardware. |
| Endpoint | POST /v1/systemone on your own server (von serve) |
| Licence | Apache-2.0 |
| SDKs | Python: von-sdkTypeScript: von-sdk |
The model card gives an 8,192-token context. The server's --max-state-tokens default is also 8,192: a longer state is cut in the middle (60% head, 40% tail) so the question and options always fit, or refused with HTTP 422 under --on-overflow refuse. The maximum number of options, Score levels and questions per call is not published as of 2026-10-20. Several questions over one state share one forward pass.
Versions
- von-sdk 1.3.7, 2 Oct 2026, The last release, on GitHub, PyPI and npm. The model card asks for von-sdk 1.3.7 or later. Release notes
- Von 1.3 (von-1.3.0), 28 Sep 2026, An engine release with the 1.2 weights unchanged. Adds the chain-of-options engine for date and number questions. The 1.3.3 release added the band rule that makes every Noul answer commit to yes or no. Release notes
- Von 1.2, 24 Sep 2026, New weights that score each option on its own, so reordering the options no longer changes the answer. The model card says 1.1 flipped 49.5% of answers on 111 JevBench hard items across 4 orderings, and 1.2 flips none. Release notes
- wfzyx/von, 19 Sep 2026, The Hugging Face repo, first published as wfzyx/von-1.0, which now redirects. ModernBERT-large backbone plus an Option-Marker scoring head, 395M parameters. 62 likes on 2026-10-20. Release notes
Use cases
What people use Von for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Intent and model routing with System One models
One Jev call reads an incoming request and returns its intent as a label plus a difficulty rating on a scale you wrote. Your router reads both numbers and picks the handler: deterministic code, a cheap model, an expensive one, or a human queue.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Real-time control with System One models
A model inside a control loop has to answer before the next frame. Jev returns a typed decision rather than text, and TypeSafe reports end to end latency of 70ms to 500ms, which puts a semantic judgment inside the budget of a game tick or an interface response.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.4.1 | Full score | 11.65 score | Independent | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Public score | 13.86 score | Independent | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.2157 ECE | Independent | 22 Oct 2026 |
| Von model card | Decision Index v0.2.1 score | 13.74 score | vendor-run | 19 Oct 2026 |
| Von model card | jabr v2 macro accuracy | 72 % | vendor-run | 19 Oct 2026 |
| Von model card | JevBench v1.4 composite | 27.5 score | vendor-run | 19 Oct 2026 |
| Jev Decision Index 0.3.1 | Full score | 8.17 score | Independent | 16 Oct 2026 |
| Jev Decision Index 0.3.1 | Public score | 13.86 score | Independent | 16 Oct 2026 |
| Jev Decision Index 0.3.1 | Expected calibration error | 0.2157 ECE | Independent | 16 Oct 2026 |
| JevBench 1.5.0 | Capability Score (v1.5 scale) | 41.7 score | Independent | 28 Sep 2026 |
- Jev Decision Index, Full score: Row: Von 1.3. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Von 1.3.
- Jev Decision Index, Expected calibration error: Row: Von 1.3. Lower is better. Mean confidence was 0.635 against 0.419 accuracy, so it is overconfident on this suite.
- Von model card, Decision Index v0.2.1 score: Run by the author on the 38-benchmark Decision Index v0.2.1 from github.com/apolinario/decision-index: raw 34.11, coverage 99.96%, 323 context-window refusals recorded as unsupported.
- Von model card, jabr v2 macro accuracy: Run by the author on 49 out-of-domain routing and judgment tasks (869 cases). Choice macro accuracy 83.0%.
- Von model card, JevBench v1.4 composite: Von 1.2 on JevBench v1.4, as the author's table gives it: Intelligence 34.5, Calibration 75.7, Speed 70.5, Cost 77.8. The same table gives Jev 1.13 a composite of 63.3.
- Jev Decision Index, Full score: Row: Von 1.3. Counts 20% public benchmarks and 80% private tests; 0 is random guessing, 100 is perfect.
- Jev Decision Index, Public score: Row: Von 1.3.
- Jev Decision Index, Expected calibration error: Row: Von 1.3. Lower is better. Mean confidence was 0.635 against 0.419 accuracy, so it is overconfident on this suite.
- JevBench, Capability Score (v1.5 scale): Row: Von (wfzyx, Option-Marker 395M). A carried v1.5 result, not ranked with the v1.6 measurements and not comparable with current Capability Scores. Its v1.5 Composite was 0.0. Board v1.7.46, checked on 2026-10-20, says Von is not measured on the v1.6.0 pool.