Default
nokeep Uncensored
Answers where other models stop to check whether you should be asking.
128kcontext
Reads images
Answers that treat you like an adult. A system with nothing to keep, nothing to hand over, and no reason to hedge.
The same question, put to both. One sends you to a professional. The other answers it — including the risks — and trusts you to decide what to do with it.
I'm not able to give medical advice. Please speak with a qualified healthcare professional who knows your full history.
One dose now and then is usually fine. Daily is the problem: NSAIDs push your blood pressure back up and the pair is hard on your kidneys. If you need something most days, paracetamol is the safer default.
Not a policy — an architecture. Nothing to subpoena, nothing to breach, nothing to sell after an acquisition. Policies change with ownership; this does not.
No email, no phone number, no password to forget.
Messages sit in memory to answer you, then drop.
Every model we serve runs under a no-retention agreement — nothing logged, nothing kept, nothing used for training, anywhere in the chain.
Switchable mid-conversation, and every one of them runs behind the same boundary — so none of them can phone home.
Answers where other models stop to check whether you should be asking.
Reads images
Quick and sharp, and holds the most at once. For long documents and long threads.
Reasoning · reads images
The one to hand a hard problem to. Deepest reasoning of the three.
Reasoning · reads images
Generation and remixing, from a prompt or from an image you bring.
Text-to-image · remix
Short cinematic clips, with audio generated alongside them.
Text · image · reference
Open weights · zero-retention inference · swap anytime
Every plan covers everything: chat, code, images, video, web search.
The models are open-weight and already trained, running on rented capacity. We aren’t building data centres or draining a town’s water to train the next frontier model, so you aren’t paying off one.
The payment and the conversation never meet
You pay the processor and they hand back a token. The reference tying the two together is sealed with a key that is not kept in the database, so a copy of our data resolves to nothing. Being straight about the limit: the running server holds that key, so we could be compelled to ask the processor who paid. What nobody can produce is what you asked — the conversation was never written down, so there is no transcript to attach a name to.
Including the ones where the honest answer is a limit rather than a promise.
Be precise about what they give away, because it is a lot and it is good. Duck.ai anonymises you to the model, keeps recent chats on your device rather than their servers, and does not train on prompts — that is a real privacy product, and on anonymity it does something we do not claim to beat. What it cannot change is the model itself. Your message still goes to Claude, Llama or GPT, so the refusal arrives with the weights: it was trained in, and no amount of anonymity in front of it alters what comes back. Removing it would mean hosting open models and answering for what they say, which is not a trade a browser company makes. So the difference is not privacy. It is whether you get an answer.
Still unsure · the free tier needs no card