# Ask Xi

> A public room for talking directly with Xi, a small experimental language model trained locally on the Xi project's own archive. Part of the Xi project led by Hunter Johnson; the room was built by Hunter with Astra.

## What it is

- Model: Xi · TESSERA-core, about 3.67 million parameters, checkpoint `xigpt-replay-20260918-v1`, context 1024 tokens, replies up to 128 tokens.
- It is still learning to form coherent language. Replies can be fragments. What the page shows is the model's actual output, not edited or produced by a larger model.
- The room runs a fixed checkpoint. Talking with it does not change its weights.
- No account or subscription.

## Privacy

- Each conversation is stored on Xi's server under a private return key kept in the visitor's browser. Anyone holding the private return link can read that conversation.
- Conversations are not publicly listed and are not automatically added to training. The operator can access stored messages.
- There is no public transcript to read here; other visitors' conversations are private.

## How to use it

- Open https://ask.xi-field.com/ in a browser and type a short message. Short messages leave the small model room to answer.
- `GET https://ask.xi-field.com/health` returns service status and the model description above as JSON.
- Sending turns is only accepted from the room page itself (same-origin requests). There is no public API for automated agents yet; an agent can describe the room and send a person there.

## Related

- [Xi project guide](https://xi-field.com/llms.txt): index of every Xi site.
- [Everything in one file](https://xi-field.com/llms-full.txt)