Project profile
Von
A 395M ModernBERT decision model with Python and TypeScript clients and a /v1/systemone server.
Last verified: 2026-09-30
Quick facts
| Project | Von |
|---|---|
| Creator | wfzyx |
| Type | Decision Model |
| Language | Python |
| License | Apache-2.0 |
| Uses Jev for | Local /v1/systemone server |
| Open source | Yes |
| Status | Active |
What it is
Von answers Choice, Noul, and Score over text or JSON without generating tokens. The README says it is a 395M ModernBERT encoder, served on CPU with OpenVINO as well as CUDA, ROCm, and Apple MPS, and that von serve speaks the TypeSafe wire protocol.
Why it matters
It is a small encoder with its own SDK and an explicit English-only limit, which is the sort of constraint a migration note has to keep.
Relationship to Jev
| Hosted Jev | Not called. |
|---|---|
| Wire shape | README: von serve exposes /v1/systemone. The repository description calls it a local drop-in. |
| Clients | von-sdk on PyPI and npm. The README also shows decide, judge, rate, and system_one helpers. |
| Language | README: English only. Other languages get token matching whose confidence it says the model has not earned. |
Architecture
- Text or JSON state
- 395M encoder
- One forward pass
- /v1/systemone response
Key features
- Several questions over one state are one forward pass, according to the README.
- Weights are fetched from Hugging Face on first use. The README says that download is about 3 GB.
- The README's curl example names the model von-1.3.0.
Performance and claims
The GitHub repository description says Von is a sub-15 ms, non-autoregressive, local drop-in alternative to TypeSafe Jev.
Jev AI Hub read the README opening and the repository record on 2026-09-30. It has not timed a decision or compared answers with hosted Jev.
The useful documented constraints are the 395M encoder, the English-only note, and the wire path. The latency line is a repository-description claim until someone measures it on their own machine.
Limitations
- The README says judge without criteria is the weakest path.
- English only.
- The repository description says sub-15 ms. The README opening reviewed here states the model size and the CPU path, not that latency figure.
Code and repository
GitHub snapshot captured 2026-09-30: 779 stars, 57 forks, 2 open issues. Last push 2026-09-30. Counts change; this is not a live lookup.
Related Jev concepts
Related guides
Related projects
Sources
- Von READMEwfzyx · accessed 2026-09-30 · github
- wfzyx/von repository metadataGitHub · accessed 2026-09-30 · third-party