How to Use Carpathian's Free AI Chat
Carpathian AI is free, needs no account, and runs models on hardware we own in the United States. Generate images, upload documents, all through Carpathian AI.
Carpathian AI is free, needs no account, and runs models on hardware we own in the United States. Generate images, upload documents, all through Carpathian AI.
How branded short links work, how to set one up on your own domain with a TXT and a CNAME record, which redirect to use, and how to pick slugs people can retype
Training resolution, aspect ratio, duplicated limbs, and why generate-then-upscale beats generating huge in one pass.
Per-image, credit, and subscription pricing compared on how they behave, plus how to estimate spend from the number of attempts a usable image takes.
A working method for writing an AI image prompt: what to put first, which words carry weight, what negatives are for, and how to iterate without guessing.
Carpathian's AI chat is a free AI running models on hardware we own, hosten in the United States. No sign up, no message limit, no token limit, no card. We want to offer it with no trial that runs out on day three, no daily cap that resets at midnight, and no counter in the corner telling you how many messages you have left. Chat for as long as you want, come back tomorrow, do it again.
If you later want to build on the same models, they're behind an OpenAI-compatible API and you already know how they behave.
It's actually unlimited. Here's the fine print anyway, because every "free unlimited AI chat" on the internet says that and most of them are lying.
We do not cap your messages. We do not meter your tokens. We do not cut you off after an hour. Replies aren't capped either, so a long answer runs to the end instead of stopping mid-sentence.
What we do have is a burst rate limit on requests per minute from one IP address. That exists because of scripts hammering the endpoint, not because of you. If you're typing with your hands you will never see it.
The other thing is a one-time agreement before your first message, where you confirm you're 18 or older and accept the Terms and the Acceptable Use Policy. We store that per browser, so you'll see it again on a different machine, or if we change those documents.
That's it. Two things, and neither one is a meter.
No. You can use the whole thing signed out: unlimited chat, turn thinking on, attach files, paste in a picture and ask about it, vote on answers, ask it questions about Carpathian.
What you give up signed out is history. Your transcript lives in your browser, which means a refresh is fine but clearing your browser isn't, and there's no list of old threads to go back to. That's usually the thing that pushes people into signing up.
Unlimited means unmetered. You get as many messages as you want, for free, forever, and that part is real.
Unrestricted is a different question and the answer is no. Every model here is a model with its own training behind it, and each one draws its own lines about what it will and won't do. If one model won't go somewhere, another one in the lineup might, and switching is one click. But if you came here looking for a model with no guardrails at all, that isn't what this is.
Fine-tuning teaches behavior, RAG supplies knowledge, prompting shapes both. What the published comparisons found, and how to pick without burning a GPU budget.
We do screen content against the Acceptable Use Policy. That screening is written to leave normal use alone, and that explicitly includes security research and creative writing that goes to dark places.
The picker at the top of the chat splits the lineup into three groups. Free models are open to anyone, signed in or not. The ones under "With an account" cost nothing once you sign in. Paid models bill per token at the rate printed next to them, and those need a payment method.
Each model also has a grade, Lite, Medium or Large, and the header next to the picker tells you which grade you're on and links you to something stronger if there's something stronger.
Start on a free model and just try your question. The free ones are the smaller models, and small is plenty for a rewrite, a summary, a quick script, or anything with a known answer. You'll know when to move up. The answer comes back thin, or the problem has four steps in it, or you're in some corner of a subject where the small model starts making things up. Switch in the picker, or type /model and a name if you'd rather not touch the mouse.
All the rates are on the AI model pricing page. That page reads from the same catalog the dashboard bills against, so the number you see there is the number you get charged. (We built it that way on purpose. Pricing pages drift.) If you want the background on why open weights matter at all, Open Source AI Models covers it.
On models that support it, there are two pickers above the message box. Instant or Thinking, and then Light, Medium or Deep. Instant just answers. Thinking lets the model work through the problem first, and the level controls how long it's allowed to sit there. Whatever you pick sticks to that thread.
Leave it on Instant most of the time. We ran the same prompt both ways on one of our hosted models: 2 seconds and 164 tokens on Instant, 39 seconds and 3,147 tokens with thinking turned on. That's a real difference and most questions don't earn it.
Turn it on for a plan, a proof, code with more than one moving part, or when you catch yourself thinking "give me a second on that." Deep is for the ones you'd normally sleep on first.
Paperclip, or drag the file onto the window. The panel that opens sorts things by how they relate to the thread: what's going out with your next message, what's attached to the whole conversation, your library of past uploads, and anything the assistant has written for you. You decide whether a file rides one message or stays for the whole conversation, and you can pull something out of the library instead of uploading it a second time.
PDFs, plain text, Markdown, images. We take text out of the file's text layer when there is one and run OCR when there isn't, so a scanned contract or a photo of a whiteboard comes through as text the model can actually read. That makes this a free unlimited chat with PDF too, with no daily cap on how many you ask about.
Pictures work the same way. Send several in one message and ask about any of them ten messages later. If the model you picked can't see images, it gets a written description instead, so you don't lose your model over one screenshot.
Signed out, a conversation holds four files. Signed in, they go into your account's AI files, and even an upload that failed to process can still be opened and read.
Describe what you want, and a model with tool access will hand the request off to the image model itself. Or type /imagine and your prompt, which works on every model and skips the chat model entirely. That second one is faster. Put a size on the end if you care: /imagine a lighthouse at dusk 1024x768. Default is 512x512, sides have to be multiples of 16, and one prompt gives you up to four images.
They show up under the reply with the prompt next to them. While it's running you get a tile per image, a clock, and a progress bar once the model starts reporting steps.
This is the one part that isn't unlimited. Images need a free account, every account gets an image allowance on a rolling window, and past that each image is priced per call. The allowance and the per-image price are both on the pricing page. A generation that fails doesn't spend anything. When you do run out, the chat tells you when the next one frees up, or where to put a card if there isn't one on file.
Ask for a report, a plan, a letter, a CSV, and say you want it saved. A model with tool access will write the actual file instead of dumping it into the reply for you to copy out.
It appears under the reply with a filename. Click it and it downloads. It stays with the conversation until you delete the conversation, and it also sits in your AI files next to your uploads and images. Text formats only: Markdown, plain text, CSV, JSON, HTML. 2 MB ceiling.
There's a ring next to the message box that fills up as you eat into the model's context window. Hover it if you want the actual numbers. When it's getting tight, click it or type /compact, and we'll squash the thread down into a summary and keep going from there. /new starts something fresh. /clear deletes what's in front of you.
Signed in, the reply doesn't belong to the tab that asked for it. It runs on our server. Close the tab, switch threads, reload the page, and the reply keeps writing itself. Come back and it's all there from the first word, and a thread that finished while you were gone gets a dot on it until you look. On a slow one we'll offer you a browser notification so it can find you anywhere on the site. Hit stop and you keep whatever the model already said.
You don't have to go find a support bot. Every turn gets checked against our own documentation, and when your question matches something in there, the answer comes out of the docs instead of out of the model's memory. Ask "do you charge for this chat?" or "which regions do you have?" and you get the documented answer. Ask about anything else and the documentation stays out of the way.
It also knows what day it is, in your timezone if you've set one, so "next Tuesday" lands where you meant it.
Create an account from the chat and it drops you right back into it.
We store conversations so you can come back to them, and message content is encrypted at rest. We also log the IP address and browser a conversation came from, which we need for handling abuse, and only authorized staff can see those. Signed-out conversations get stored the same way.
Prompts, conversations and generated content get used to run and secure the service, and to develop, train, evaluate and improve the models. That's Section 21 of the Terms of Service. We don't sell any of it, rent it, or license it out, and none of it touches advertising. Your name, email and account details are never used for training.
Deleting a conversation pulls it out of your view and out of our active systems. Security records can outlive that, and copies can sit in backups for a while. Generated images get removed after a retention window whether you delete them or not, so download the ones you want to keep.
Same things every language model is bad at:
It says wrong things in exactly the same confident voice it says right things. It's weaker on niche material and on anything recent than it is on well-covered ground. It can't check its own work, so it will hand you code that reads perfectly and doesn't compile. Turning thinking on and moving up to a bigger model lowers how often that happens. It does not fix it. Anywhere being wrong would actually cost you something, go check the answer against something that isn't a model.
When a model doesn't come back at all, the chat says so in one sentence, tells you it looks offline, and gives you a link to report it with the failure already filled in (you don't have to describe anything). Pick a different model and the notice clears.
When the thing you've been poking at turns into something you want to ship. A browser tab can't be called from code, can't be given a rate limit, and can't be pointed at a budget.
Our inference API follows the OpenAI chat completions format, so anything written against the OpenAI SDK moves over on a base URL and a key, and the same models answer on the other side. Charged per token. Every request logged with its token counts and response time. Each key carries its own rate limit, IP allowlist and token budget, so you can hand one out without handing out the whole account. AI Inference API: OpenAI-Compatible Endpoints has the calls, and Carpathian AI covers the hosting side.
Doing that doesn't change anything about the chat. It stays open and it stays free.
Is there a free AI chat with unlimited messages? Yes. This one. No message cap, no daily reset, no token meter, and no card required.
Can I use it without signing up or logging in? Yes. Signed out you get unlimited chat, model switching, thinking mode, and file and image uploads. The only thing an account adds is storage: history, agents, image generation, and replies that keep running after you close the tab.
Is there a token limit?
No. We don't meter tokens on the free models and we don't cap reply length. The only ceiling is the model's own context window, and when you get close to it you can type /compact to summarize the thread and keep going in the same conversation.
Is it free forever or just a trial? Free, with no trial window. The free models stay free whether you make an account or not.
Is it an unrestricted or uncensored AI? No. Unlimited and unrestricted aren't the same thing. Usage is unmetered, but each open source model in the lineup has its own limits about what it will answer, and content is screened against our Acceptable Use Policy.
How many files can I upload? Four per conversation signed out. Signed in, uploads go to your account's AI files and stay there. PDFs, text, Markdown and images, with OCR for anything scanned.
Are images unlimited too? No, images are the exception. They need a free account and come with a rolling allowance, and past that they're priced per call on the pricing page.
Where does it run? On hardware Carpathian owns, in the United States.