Jev

Jev is a hosted decision model API from TypeSafe. You send it a state (text or JSON) and a set of typed questions, and it returns typed answers with probabilities rather than generated text. Its question types are choice (pick one of several labels), score (pick a level on an ordinal scale) and noul (a yes/no probability). The HTTP interface is POST /v1/systemone.

Jev is closed: the weights are not published and it is available only as TypeSafe's API. Its published price is $0.042 per million tokens.

Jev and Laya

Laya is an open-weights (Apache-2.0) model from Convai Innovations whose output is schema-compatible with Jev. The Laya package ships laya-serve, which exposes the same /v1/systemone request and response shape, so a client written against Jev can point its base URL at a Laya server and keep working. Laya Studio hosts Laya behind that protocol. Laya Studio is independent and not affiliated with TypeSafe or with Convai Innovations.

Published comparison points

All Jev figures below are third-party or published, not measured by Laya's maintainers or by us:

• Latency: 236–276 ms p50, measured by AbdelStark and nibzard. Laya reports 32.8–39.5 ms per question on a T4.
• High-cardinality labels: Jev supports up to 255 options and scores 0.870 on Banking77 (72 labels); Laya scores 0.425 (77 labels) at default settings. Jev is the better fit for very large label sets out of the box.
• typed-decisions: Jev 1.13.0 scores 0.727 accuracy and 0.580 soft accuracy; the fine-tuned Laya checkpoint scores 0.766 accuracy and 0.471 soft accuracy.
• Calibration: Jev's published ECE is 0.144 on typed-decisions and 0.246 in the model card's headline table.

See Laya vs Jev and Jev alternatives for the full comparison, and what is Jev for the background.