Free AI chat, with nothing to sign up for
Open the chat and start typing. There is no account to create, no quota counting down in the corner, and no card to put on file. The models answering are open source models we host on hardware we own in the United States, and the same models sit behind an API for the day you want to build something on them.
base_url at https://api.carpathian.ai/ai and keep the rest of your code.The public chat runs the same models the API serves.
What free means here
Free is a word that usually carries a footnote. It tends to mean a number of messages per day, or a smaller model than the one the paying tier gets, or a window that closes after two weeks and asks for a card. None of those apply here.
The chat is open to anyone who loads the page. There is no account gate in front of it, no quota behind it, and no better version of it held back for a plan. One limit does exist, and it sits in the list with everything else.
- No account, no email address, and no card to open the chat
- No message quota and no token meter counting down while you type
- No trial window that closes and asks for payment
- No paid tier holding back a better model than the one answering you
- One burst limit on requests per minute per address, which typing does not reach
- Open source models running on compute we own in the United States
- The same models reachable through an OpenAI-compatible API when you build
Signing in changes what you can reach in the dashboard. It does not change what the chat costs or how much of it you can use.
Where a message goes
Four steps, and none of them involve a third party deciding how much of the conversation you are entitled to.
Message
public chat- You typeThe message goes to Carpathian, not through a reseller sitting in front of someone else's consumer product.
- A model we host answersAn open source model, loaded on our own compute in the United States, in the same environment that serves the inference API.
- Nothing accumulates against youThere is no balance to draw down and no daily cap, so the second hour of a conversation behaves the way the first one did. A burst limit keeps one client from taking the endpoint away from everyone else, and typing does not reach it.
- You can take it furtherWhen the thing you are testing turns into something you want to ship, the same models answer to an API key instead of a browser tab.
What it is good for, and what it will get wrong
A model is a tool with a shape. These are the jobs it handles well and the ones where it will hand you something wrong with a straight face.
Drafting and rewriting
Getting a first version onto the page, then cutting it down. It is at its best when you already know roughly what you want to say and the work is in the saying. You still have to read the result and judge whether it landed. Ask it for three openings rather than one, and keep the one that sounds like you.
Reading something long
Summarizing a document, pulling the decisions out of a meeting transcript, or answering a question about a wall of text you paste in. Give it the source and ask about the source, which is the case where a model is on firm ground because the answer is in front of it.
Code you are stuck on
Explaining an error, sketching a function, or talking through an approach before you commit to it. Treat it as a fast second opinion rather than an authority. It will produce code that looks right and does not compile, and reading the output is still your job.
What it will get wrong
It states wrong answers in the same tone as right ones. It is weaker on niche and recent material than on well-covered ground, it cannot check its own claims, and it will invent a citation rather than admit it has none. Where being wrong would cost you something, check it against a source that is not a model.
Why it costs nothing
Inference is not free to serve, so the question is who absorbs the cost. For most free chat products the answer is a funding round, an advertising business, or a paid tier the free one exists to advertise.
Ours is narrower. We own the compute and the buildings it runs in, and much of that hardware came out of the e-waste stream and was rebuilt for service rather than bought new. We run models sized to the work instead of the largest thing we could load, and our research goes toward making inference cheaper to serve. Free is what that cost position looks like from the outside.
It also does something for us. The chat is the shortest path from reading about the models to using them, and a person who has already used one is a better judge of whether the API is worth their time.
When the browser tab is not enough
The chat is a place to try the models. It is not where you put them once they are load bearing, because a browser tab cannot be called from your code, given a rate limit, or pointed at a budget.
That is what the inference API is for. It follows the OpenAI chat completions format, so an application already written against the OpenAI SDK moves over on a base URL and a key. Usage is charged per token, every request is logged with its token counts and response time, and each instance carries its own rate limit, IP allowlist, and token budget.
Nothing about the chat changes when you do that: it stays open and it stays free.
Questions people ask before the first message
Is the AI chat free, or is it a trial?
Do I need an account or an email address?
Is there a daily limit or a message cap?
What is answering when I type?
Can I use the same models in my own application?
Why give the chat away when inference costs money?
What is the chat bad at?
The rest of the AI work
The models page covers what open weights are and why we run them. The efficiency page is the argument for smaller models on hardware that already exists, with the measurements attached. Veritate is the research program those measurements come out of.
Open it and ask it something.
No account, no card, no counter. When the thing you are testing turns into something you want to ship, the same models answer to an API key.