Compare the local projects
| Project | Documented scope | Decision that changes the choice |
|---|---|---|
| JEV-MLX | MLX-based local choices from application state, an utterance and dynamic candidates on Apple Silicon | You need a native Apple Silicon MLX checkpoint and an explicit no_match or abstain path |
| Jev Visual | Qwen3.5-0.8B with MLX for choice, yes/no and ordered-score questions about one image | You need image input and accept probabilities that are relative to the supplied options |
| OneJev | A proposed multimodal System One model for screenshots, photos, video and text, with a local server | You need a PyTorch serving path and media-aware requests rather than a small Apple Silicon experiment |
| Hosted Jev API | TypeSafe's remote System One endpoint and official SDKs | You prefer the documented provider and do not require local inference or local weights |
The repository names describe relationships, not equivalence. JEV-MLX and Jev Visual state that they are independent experiments; OneJev is a separate open model. None of these pages is evidence that TypeSafe's proprietary weights, training or calibration have been downloaded.
What JEV-MLX is useful for
The JEV-MLX README describes a Python API, CLI and localhost HTTP service. Its request combines the current state, a user utterance and allowed choices, then returns a selected business ID or an explicit rejection. State-version checks and single-use execution authorization are application safeguards; they do not make the model's semantic decision correct.
Its documented measurements are tied to named checkpoints, an Apple M2 Max environment and fictional English cases. The project itself says candidate scores are not calibrated probabilities of correctness and that version 0.1 is experimental. Treat those measurements as the author's evaluation record, not a JevHunt benchmark or a general performance promise.
What Jev Visual and OneJev add
Jev Visual's pinned README documents a local browser UI, CLI and HTTP API for one image. It reuses shared multimodal context and directly scores candidate outputs. Its own limitations say that only the Qwen3.5 adapter is verified and that option probabilities are relative to the options supplied.
OneJev's pinned README documents a /v1/systemone server with optional image and video inputs and a qev serve quick start. Its published tables and latency figures are project-authored results; reproduce them with the stated model, hardware and test set before using them for capacity planning.
Choose a first local test
- Choose a modality: text and bounded choices for JEV-MLX, a single image for Jev Visual, or media requests for OneJev.
- Pin the repository revision, checkpoint, runtime and numerical precision in the test record.
- Run a small labeled set with an option that means “none of these” or an abstention path.
- Separate model choice from execution: a state-version check can block a stale action even when the model selected it.
- Compare a hosted baseline only when the state, questions and scoring criteria are held constant.
For more independent models and adapters, use the local alternatives guide. For a broader scorecard covering calibration, review rate, latency and cost, see how to evaluate a decision model.