Open-Source System One Models

Hosted Jev is a closed model behind POST /v1/systemone. The fields on that request are documented, so other models can answer the same three question types: Choice, Score, and Noul.

Servers that speak the request

These READMEs document a local POST /v1/systemone. An existing client changes its base URL. The weights change with it.

Project	Stars	License	What we verified in the README
Kev	7,965	Apache-2.0	Python SDK drop-in. Sizes 0.8B, 4B, 9B, 27B on Qwen bases.
Laya	28,825	Apache-2.0	Router over a 421M English checkpoint and a 322M multilingual checkpoint. laya.serve exposes the path.
Von	779	Apache-2.0	395M ModernBERT. Python and TypeScript SDKs. README says English only.

Kev's own table puts Kev-27B within a point of Jev on development sets the README calls "new sources," and it says Jev was not run on the test half. Use that sentence with the number.

Two smaller servers showed up in the same search and do not have a full profile yet:

• khimaros/verdict (7 stars, GPL-3.0). The README says a Python server in front of llama-server reads option-label probabilities and serves the same wire shape.
• featherless-ai/simple-jev (574 stars, Apache-2.0). The README says /v1/systemone is an alias of its classifier route, and that the eval list includes a vision suite. Hosted Jev's inputs on the pages we cite are text and structured state.

Libraries on a model you already run

No new decision checkpoint. You bring an open LLM, and the library reads a distribution over labels you supply.

Project	Stars	License	What the README says
SemIf	4,591	MIT	Formerly OpenJev. Option probabilities from a frozen model. Browser demo. Not affiliated with TypeSafe.
AnyJev	974	Apache-2.0	L0 rotates options with no labels. L2 fits a small head when you have on the order of 100–300 labels.
open-alternative-jev	57	Apache-2.0	In-process package (so1) for Hugging Face and vLLM. The README says there is no drop-in HTTP API for the official SDK.

A small model trained in public

NanoJev (2,438 stars, MIT) is Qwen3-0.6B plus decision heads. The README's held-out table is maze, Snake, and two ViZDoom tasks, with hosted Jev as one of the comparison systems. That is a game controller eval. It is the wrong table to quote if you are routing support mail.

Tools that call the real API

These do not replace the model. They are still the clearest public examples of Jev inside a product loop:

• Jev Ultrafast picks a browser operation.
• jevc compiles policy files into questions and a code reducer.
• pi-jev-auto-mode blocks tool calls unless the gate can decide.
• jevmetrics scores OpenTelemetry metric metadata. The README says alpha.

What to compare before you switch

1. Wire compatibility. A documented /v1/systemone is a different test from a library with its own decide() method.
2. Language and modality. Laya's README routes non-English text to the multilingual checkpoint. Von's README says English only. Simple Jev's README mentions vision. Hosted Jev, on the official pages we use, does not document image input.
3. Calibration. Choice confidence and a Noul are only useful if the threshold means what you think. Fit that on your labels. Author Brier scores do not transfer.
4. Who can fine-tune. Open weights can be trained by the person who holds them. Official models docs say hosted Jev is not fine-tuned on customer data.

What this page leaves out

Dozens of other checkpoints appeared in the same two weeks. A directory that repeats every self-reported accuracy becomes a leaderboard this hub cannot stand behind. New profiles get added after the README is read and the GitHub record is captured. The method is the same one used for Jev Ultrafast: project claim, what we actually opened, and what a reader should not conclude.

Next: Jev GitHub or the project index.