Trained models
Own decision heads
Downloadable weights for local serving and research.
Inference engines
Reuse base weights
Read logits or broadcast KV cache to produce Jev-like outputs.
Compatibility
Match the API shape
The interface can match while the underlying quality differs.
Projects to start with
| Project | Type | Base / feature |
|---|---|---|
| Mapika/decider-2b | trained model | Qwen3.5-2B, Choice / Score / Noul |
| Laya | trained model | ModernBERT, 421M |
| NanoJev | research | Qwen3-0.6B, games and mazes |
| zhihz/openjev | local preview | Qwen3-4B, bilingual experiments |
| SemIf | engine | Native candidate-probability readout |
Which path should you pick?
- General local API: start with decider-2b.
- Mac experiments: look at SemIf, open-jev and jev-on-a-laptop.
- Bilingual research: test zhihz/openjev with your own data.
- Teaching demos: look at NanoJev and jevlike.
Shared disclaimer: These projects are not official TypeSafe AI Jev weights. Speed, accuracy and comparisons should be treated as author-reported until independently reproduced.
Model cards
Each model page records weights, base model, license, context, language, question types, calibration evidence and failure cases.