Kuyun AI · Works to try

Help · The AI models behind Kuyun

The AI models behind Kuyun

The short answer

One model writes almost everything, and a second, different model reviews what the first one produced before it becomes a finding in a project. When a big job in a project's chat carries on by itself, the reviewer also reads each new part against the sample you approved and decides whether the work goes on, and it is the model that marks a written answer on a practice paper when you ask for that. Today, in the order they are tried: for writing — your chats, your project research, summaries, and everything else Kuyun writes — DeepSeek, served by DeepInfra, in the United States; if that fails, DeepSeek, served by DeepSeek's own service, in China; and if that fails too, DeepSeek, served by Alibaba Cloud's international service; and if that fails too, Anthropic's Claude, served by Anthropic, in the United States; for reviewing — checking a project's brief, its plan and each step's work before it becomes a finding, judging whether a question really needs you and whether a big job in a project's chat goes on by itself, and, once a week, reading a dozen of the replies you received, with what you asked, to judge how well they were written for you — GLM, served by DeepInfra, in the United States; if that fails, Qwen, served by DeepInfra, in the United States; and if that fails too, Anthropic's Claude, served by Anthropic, in the United States; for drawing a picture a chat makes — the few words describing it that the chat writes — OpenAI's GPT Image, served by OpenAI, in the United States; if that fails, FLUX, served by DeepInfra, in the United States; and if that fails too, Google's Gemini, served by Google; for reading aloud — the text you press 🔊 on — OpenAI's gpt-4o speech models, served by OpenAI, in the United States; if that fails, Google's Gemini, served by Google; and if that fails too, Qwen, served by Alibaba Cloud's international service; for turning what you say into text — your recording — OpenAI's gpt-4o speech models, served by OpenAI, in the United States; if that fails, Google's Gemini, served by Google; for a live spoken conversation with a work's study partner — your voice as you speak, and what the study partner is told — Google's Gemini, served by Google; if that fails, OpenAI's realtime models, served by OpenAI, in the United States; and if that fails too, OpenAI's realtime models, served by OpenAI, in the United States; for finding works similar in meaning to a search — the words you type into Explore's search — BAAI's BGE-M3, served by DeepInfra, in the United States. If the first fails, the next answers, and so on down the line.

Why two

A reviewer that is the writer's own model inherits its blind spots. Having a different model check the work catches things the writer was never going to notice about itself. The two may come from the same company, as two of its models of different size, or from two companies. This matters most in projects, where a finding is supposed to be something you can rely on.

Choosing your own, per project

A project can be set to use different models, task by task: one for everything that writes and one for everything that checks, and, where it matters, a different one for a single task — the plan, the research steps, the reviews, the Creation Console, the viewer's chat and so on. A task without its own uses its group's, and a group without one uses the default; the two chats join no group, so only a pick made for them moves them. This is done when starting the project (Models (admin) on the New project card) or later from Models… in the project page's ⚙ menu.

Only the project's owner chooses, and only when the owner is an administrator: nearly every task sends the project's material — your files, your problem statement, your questions, and the words of the people you share it with — to the model that does it. An administrator you share your project with cannot change any of them. A project nobody has set this on uses the defaults.

A model chosen this way is tried first, and the defaults for that task stand behind it. If its company has run out of credit or is not answering, the task goes on with the default models instead of stopping, and goes back to the chosen one once its company answers again. Either way the work only ever reaches companies the terms list.

An administrator comparing models can also choose, for their own conversations only, which model writes replies to them, which turns their own recordings into text and which reads replies aloud to them (Models to test… in a viewer's ⚙ menu, shown to administrators alone). It follows them into their main chat and to every work they view; the Creation Console of a project they make keeps its own setting. Their conversation with your work then reaches the model they chose, as any viewer's conversation reaches the default one — a model of the same companies the terms list. Nothing changes for you, for your work or for anyone else who views it. A model they chose for hearing or speaking answers alone: if it fails, they are told so and no other answers in its place. The one they chose for replies is tried first, and the usual fallbacks may answer when it cannot.

A spoken conversation with a work's study partner is held live — one model that hears your voice and answers in its own over an open connection, in place of the recording, the written reply and the reading aloud (Talking instead of typing, Conversation mode). Google's Gemini live model holds it first; when that cannot open a conversation, OpenAI's gpt-realtime models do, the smaller before the larger (the list above gives the order in force). When none of them can, it is held the recorded way and you are not asked anything. Your voice goes to that model's company, one of those the terms list, for as long as the conversation is open: straight from your browser to Google, or through Kuyun to OpenAI.

An administrator can choose which live model holds their own conversations — the fourth choice in the same Models to test… — whatever everyone else is served, and nobody else's conversation changes. A model they chose answers alone: if it cannot open a conversation they are told why, and theirs is held the recorded way.

This is per project, not per account. It is worth doing when a project has a particular character: something requiring careful reasoning might deserve a stronger and more expensive model; something routine does not.

If a model you chose stops working, that project tells you rather than silently switching to another one.

What happens to what you type

Your messages, and the parts of your documents and memory relevant to the question, are sent to the model provider to produce an answer. This is how the answer gets made — there is no version of this that keeps your words on our machine.

Which company depends on the model in use. By default it is one of those listed in "The short answer" above, in that order. The companies that may ever receive your words are listed in the terms (Terms of Service and Privacy Policy); the default models are only ever chosen from among them.

Drawing pictures

When a picture would help and none can be found — an illustration for a lesson, a labelled scene, a vocabulary card — a chat can draw one: your Home chat, a project's chat, the study partner of a course shared with you, and the background work that writes a project's lessons or chapters. The chat writes a short description of the picture, and an image model draws it: OpenAI's GPT Image (gpt-image) at its quickest quality first, then FLUX on DeepInfra, then Google's Gemini image model, each trying when the one before cannot. Only the description goes to them, never your conversation.

A drawn picture is kept like a found one — in the project, or, when the chat was only yours, among your own pictures — and shows a "drawn by" credit. It costs tokens like any other model use: what the image model reports, or 1,500 tokens when it reports none. It is charged to whoever's chat it was drawn in; a lesson's picture drawn by the background work is the project owner's.

To illustrate lessons or chapters that are already written, ask the project's chat — "illustrate the lessons" — and it decides what helps: how many pictures, of what, and where each goes, beside the part it shows. The rest of each lesson is left exactly as it was, and no one's place in the course moves. Drawn pictures carry no words: image models misspell them, so the word a picture teaches is in its caption.

Reading aloud and listening

When Kuyun reads something to you — a 🔊 button, a dialogue in a lesson — the text goes to OpenAI's gpt-4o speech model first, then, when it cannot answer, to Google's Gemini, then to Alibaba's Qwen. Each has its own voices, a woman's and a man's for every speaker, and one of them reads a whole clip, so a conversation never changes voices halfway.

When you speak instead of typing, your recording goes to OpenAI's gpt-4o transcription model, and to Google's Gemini when it cannot answer. Only the recording goes, never the conversation around it.

Searching by meaning

What you type into Explore's search also goes to BGE-M3, a model from BAAI served by DeepInfra, which turns it into a list of numbers that stands for its meaning, in any of Kuyun's three languages. Each published work's public page is turned into numbers the same way, so search can find a work that means what you asked for even when none of your words are in it. Only the words in the search box go, never anything else of yours, and they are not kept (Explore).

Speed and cost

The default writing model was chosen for being fast and inexpensive rather than the most capable available; the default reviewer is a more capable model. If a particular project needs more, an administrator can change its models; that is what the setting is for.

What it cannot do

Whichever model is chosen, Kuyun can only read, think and write. It cannot send email as you, spend money, or contact anyone. Changing the model does not change that.

This is a page of Kuyun's own handbook, the one its assistant answers from — ask it in the app and it will say the same.