Pricing
llamay runs GGUF models on your machine and costs nothing to do it — no account, no card, no ceiling, and nothing leaving the machine. What the plans below buy is the part that cannot happen on one computer: a conversation that picks up where you left it on another, and models too large to run on yours.
Loading prices…
Billed monthly, or yearly where a yearly price is offered, and cancellable in a click. llamay never sees or stores a card: your subscription is billed by Polar, llamay's merchant of record, which is the name that appears on your statement.
One thing, mostly. It is worth saying plainly rather than putting it in a table as a tick.
Every other engine stores your conversation as text. Reopen it and the model reads the whole thing again before it can answer — thirty seconds of a long thread, every time, on every machine.
llamay stores the computed state instead: 140–450 MB that resumes in under a second. Syncing that state is what Pro buys. The conversations themselves sync on the free plan.
Hosted chat models run on Cloudflare Workers AI through llamay’s service, and most of them run on any plan, paid for out of that plan’s token allowance.
The ones that cost over a dollar a million tokens to run — GLM 5, Kimi K2, DeepSeek V4 Pro and the rest — are the exception. Those are what Plus buys. That threshold is the rule the service applies, so which models fall either side of it follows the price list rather than this page.
A conversation is a few KB. Its attention state is 140–450 MB, so a working set of them is measured in tens of gigabytes and a library of them in hundreds.
Storage is sold separately from the ladder, because an account on Pro that has filled its 10 GB does not want Plus — it wants somewhere to put the next conversation.
Generated from the same table the service enforces, so this page cannot promise a limit the service does not honour.
On top of whichever plan you are on, and changing nothing else about it.
A seat covers commercial use, and is a signed file rather than an account.
What a hosted call cost, and not what it said.
The text of hosted prompts and answers. The service holds them for the length of the request and writes them nowhere, and every model call asks Cloudflare’s AI Gateway to log metadata only. The one exception is the summary of a conversation too long for its model, kept for thirty days.
Your account, a usage ledger (model, tokens, cost and time for the last 500 entries), hashes of your sign-in tokens and API keys — never the keys — and what your apps sync. Download or delete all of it from your account page.
Cloudflare Workers AI, whose terms rule out training on customer content. A local model never leaves your machine at all. Everything kept, and for how long.
To run models on your own machine from the command line, no — install it and go. The desktop app asks you to sign in, because everything it adds over the CLI is an account: sync, the hosted models, your plan. Signing in is free.
You drop to the free plan. Nothing is deleted and nothing is held hostage: your conversations keep syncing, your local models keep running, and the attention state already in your storage stays there and stays downloadable. What stops is new state syncing up. One limit: if what you stored comes to more than twice the free plan’s room, new conversations wait until you delete some.
In a click, from the billing portal in Settings, and it takes effect at the end of the period you have paid for. There is no call to make and no retention flow to sit through.
No. Local models never leave your machine at all. Hosted requests go to Cloudflare Workers AI, which does not train on them, and their text is not kept (above) — and the bytes in your storage are under a prefix only your account can reach.
Only hosted inference. Anything running on your own hardware is uncounted and unlimited, which is most of what llamay is for — the allowance is for the models running on ours.
Where a yearly price exists it is on the card above, with the saving worked out from the two live figures rather than written down. Monthly is always available and is the default: it needs no promise about what llamay will be in twelve months.
It is not a trial and it does not expire. Upgrade the day something in it stops being enough, and not before.