Free open-source chatbot demo where every word is picked by the Jev choice model from a 1,772-word vocabulary.
JevGPT chat landing page
1 / 2Overview
What is JevGPT?
A chatbot built on a model that can't write
JevGPT is a chatbot built on a model that can't write — every reply word is picked by the Jev choice model from a vocabulary of about 1,772 words. It re-ranks among the top-12 continuations by default, and a look-inside view shows word probabilities with a heatmap.
You can use it at curata.com/jevgpt with your own TypeSafe API key (about a cent per 20-word reply) or run it locally from the open-source GitHub repo. It is an experimental research demo: short replies, a small vocabulary, and 70-500ms per word. Free to try with a demo video available.
- Every reply word picked by the Jev choice model from ~1,772 words
- Re-ranking among top-12 continuations by default
- Look-inside view: word probabilities with heatmap
- Temperature, top-k, top-p, repetition penalty, history toggle, streaming
- Self-hostable; API key stored in an httpOnly SameSite=strict cookie
Platforms and languages
JevGPT Availability
Platforms
Languages
Capabilities
JevGPT Key Features
Word-by-word choice model
Every reply word is picked by the Jev choice model from a 1,772-word vocabulary — the model literally can't write outside it.
Top-12 re-ranking
Re-ranks among the top-12 continuations by default, improving fluency and relevance scores.
Look-inside probabilities
Inspect word probabilities and a heatmap to see how each reply was formed.
Sampling controls
Temperature, top-k, top-p, repetition penalty, history toggle, plus streaming with Stop.
Self-hostable and open source
Run it locally from the GitHub repo, or use the hosted demo with your own API key stored in an httpOnly cookie.
Best for
Who uses JevGPT?
AI/ML enthusiasts, developers, and researchers
Experiment with an open-source chatbot that picks every word from a tiny vocabulary, and inspect its word probabilities
Plans and access
JevGPT Pricing
Free
Free and open source; about a cent per 20-word reply via your own API key.
Common questions
JevGPT FAQs
How does JevGPT work?
14 word groups of up to 255 words each; each reply word is chosen by P(group) times P(word|group) from the ~1,772-word vocabulary.
Does it actually work?
As a research demo: re-ranking lifted measured fluency from 2.56 to 3.17 and relevance from 2.90 to 3.31.
How much does it cost?
The demo is free to try; with your own API key a 20-word reply costs about a cent at $0.042 per million input tokens.
What are its limitations?
A vocabulary of about 1,800 words, short replies, and 2 requests per word at 70-500ms each — it is experimental by design.
How is my API key stored?
In an httpOnly, SameSite=strict cookie, so page JavaScript cannot read it.
Reviews
See what the community thinks and share your experience.
—
Based on 0 ratings
Rating distribution
Community reviews
No reviews
No written reviews yet.
