JevModel for AI text classification and structured decisions

JevModel is a browser workbench and API for AI text classification and bounded decisions. It accepts evidence as text or JSON and returns an answer your application can inspect: a category, a rubric score or a yes probability. It is aimed at developers and teams working on support routing, moderation, lead qualification and similar repeatable decisions.

What does JevModel actually do?

Imagine a support inbox that needs to distinguish billing questions from technical problems. You supply the message and define the possible queues; JevModel asks the Jev model to choose between them. Your software then decides what to do with the result.

Three question types support different jobs. Choice selects one supplied label, Score evaluates an ordered rubric and may return a fractional value, and Noul gives the probability that a proposition is true. This is useful when the answer space is known; it does not write the customer’s reply.

JevModel is independent of TypeSafe AI, which develops Jev. Its live requests call the pinned typesafe/jev-1.13 model through OpenRouter. Illustrative homepage panels and offline examples are not measured model results.

Test a support-routing decision before integrating

1. Open Playground and choose the support-routing template.
2. Replace its sample state with a short, non-sensitive message such as a duplicate-charge complaint.
3. Review the question and category definitions. Keep billing and technical labels distinct and allow a review outcome for ambiguous messages.
4. Press Decide, then inspect the chosen label and probability.
5. Repeat with clear and mixed cases whose expected labels you know. Set automation thresholds from those tests.

A high model probability is not a measured accuracy rate. Review uncertain outcomes before connecting a result to a consequential action.

From individual texts to batches and branching flows

Browser Batch reads CSV, TXT or JSONL, applies a fixed question set to each row and exports results as CSV. Files are capped at 2 MB; free accounts upload up to 50 rows and any purchase raises that to 1,000. Export before leaving: processing depends on the browser session, rather than a durable asynchronous job.

Decision Flows connect questions to branches and outcomes, with an uncertain-answer route for human review. Only visited question nodes run. A flow allows up to eight question nodes and 24 total nodes, without cycles. It returns routing data; your application still creates tickets or sends notifications. If a later node fails, earlier successful paid nodes remain charged.

Use the REST API from your server

After email verification, create a JevModel key in Console and send state plus questions to POST /api/v1/systemone. Keep the key server-side. Check both HTTP status and the response code, and preserve request IDs. There are no idempotency keys, so check history and balance before retrying a lost response. The API reference explains the response envelope and errors; the public clients use ordinary HTTP rather than a JevModel SDK.

Frequently asked questions

Is JevModel free, and how is extra usage charged?

Checked 28 October 2026 on the public USD pricing page: the demo allows three runs per UTC day without signup. Verified accounts receive 100,000 tokens once and can save one flow. Further usage draws from a shared input-token balance; output is free. One-time packs start at US$4.90 for 8 million tokens. Packs do not expire, and unused subscription tokens roll over. Site capacity also applies. Check current plans and limits; do not read the trial as a recurring daily token grant.

Can it classify images or generate text?

The interface accepts text or JSON, not native image or audio input. Normal requests support one to eight questions; both serialized state and the full validated request must fit within 8,000 characters. Use another tool for OCR, writing, local inference or fine-tuning.

Are submitted texts stored?

Inputs pass through JevModel and OpenRouter to the model provider. Signed-in history contains at most the latest 100 successful runs within 30 days, with controls to delete individual or all saved runs. Account and security records are separate. Avoid secrets and sensitive personal records, and review the privacy policy before uploading.

Official resources and support

• Runnable GitHub examples: curl, Node.js and Python clients; offline fixtures are illustrative.
• Account settings and support: manage account issues and request help with stored data.