Comparison
Laya model vs Jev API
Laya and Jev answer the same kind of question. A state goes in. A Noul, Choice, or Score comes back with probabilities. Laya is open weights you run. This site is a hosted API for Jev.
This page is not the Laya project and not the TypeSafe console. Convai Innovations publishes Laya. TypeSafe builds Jev. Keys, credits, and support on jev-api.org come from this site.
Updated
What the Laya model is
Laya is an open-source decision model from Convai Innovations, under the Apache 2.0 license. Install it with pip install laya, or load the checkpoints from convaiinnovations/laya on Hugging Face. It does not write text.
One call takes a state and typed questions. choice picks one option you named. score places the state on an ordered scale you wrote. noul returns the probability that a statement is true. The English checkpoint is a 421M ModernBERT model with a 512-token budget. A multilingual checkpoint covers 100+ languages. Laya's model card says a routed call is about 33 ms on a T4 GPU.
The same card treats the base checkpoints as a starting point. On Laya's typed-decisions benchmark, the base English checkpoint scores 0.362 accuracy and the fine-tuned laya-typed-decisions checkpoint scores 0.766. Laya measured both. That pair is not a comparison with Jev.
Three ways to run the decision
Laya's model card compares Laya with TypeSafe's Jev API. This site is a third option, with tighter limits. Any Jev accuracy number quoted below comes from that card. Laya says those Jev numbers are third-party published and were not measured in the same run.
| Item | Laya | TypeSafe Jev | Jev API here |
|---|---|---|---|
| Who runs it | You, on your GPU or CPU | TypeSafe | This site |
| Weights | Apache 2.0, public checkpoints | Closed | Closed. You call jev-1.13 |
| Answer types | choice, score, noul | noul, choice, score | noul, choice, score |
| Choice options | You set the list. Wide lists need a larger head budget. | Up to 255 | 2 to 8 |
| State size | 512 tokens on the English checkpoint. The multilingual checkpoint can read longer text. | Documented on typesafe.ai | Up to 60,000 characters |
| Price | Your hardware | Per token on typesafe.ai | 1 credit per successful call. About 100 credits correspond to 1 USD before tax and local pricing. |
| Fine-tune | Yes, on your own labels | Weights are not public | No training on this endpoint |
When to run Laya yourself
Run Laya when the model has to stay on your machines, when you can fine-tune on your own labeled decisions, or when you already have a GPU and do not want a per-call fee. Fit calibration temperatures on your data before you treat a probability as a cutoff. Laya's card says the raw checkpoints are over-confident until you do that.
When to call this Jev API
Call this API when a Choice fits in 2 to 8 options and you want a key and a POST today. You do not install a model or hold a GPU — the first key works right away. A successful call costs 1 credit whether it carries one question or six. Failed calls are free. The path is POST /api/v1/decisions. The model id is jev-1.13.
Choice questions with many labels
Laya's card says TypeSafe Jev is the better fit when one Choice has more than about 20 options. On Banking77 they report 0.870 for Jev, measured on 72 labels, and 0.425 for Laya at the default token budget, measured on 77 labels. This endpoint cannot take that question. Choice stops at 8 options. For dozens of labels in one question, use TypeSafe's API at api.typesafe.ai, or split the labels into a short list for this endpoint or for a self-hosted Laya call with a larger head budget.
Laya vs Jev FAQ
Is the Laya model a chatbot?
No. It returns a typed decision and probabilities. It does not write a paragraph.
Is jev-api.org the official Laya or TypeSafe site?
No. Convai Innovations publishes Laya. TypeSafe builds Jev. This site sells independent access to Jev.
Can I send the same questions to both?
The three types match: noul, choice, and score. The HTTP path does not. Laya's server uses POST /v1/systemone, the path TypeSafe documents. This site uses POST /api/v1/decisions, with 1 to 6 questions and 2 to 8 Choice options. TypeSafe documents up to 255 Choice options. On Laya, the option list is the one you send when you run the model.
Do I have to fine-tune Laya first?
Laya says the shipped checkpoints work zero-shot, and that fine-tuning on your domain is where accuracy moves. On their published typed-decisions set, that gap is 0.362 to 0.766. Check the cutoff on your own labels before you lock it.
Which one is more accurate?
This site has not run a shared test, so there is no single winner here. Laya's card reports a higher hard-label score for its fine-tuned checkpoint on that one benchmark, and a higher score for TypeSafe Jev when a Choice has many options. Read the card, then test the questions you actually send.
Send a decision on this API
Open the playground, or follow the docs and call the endpoint with your own key. For the open weights, use the Laya project.
Figures on this page that describe Laya, or that compare Laya with TypeSafe Jev, come from the convaiinnovations/laya model card, checked 29 September 2026. Open the model card