Playground

A real language model running on your own graphics card — no key, no server, nothing leaves this tab.

This model is small, and honestly a bit stupid

It has about 3 billion parameters — roughly a hundred times fewer than the model behind ChatGPT. It forgets fast, contradicts itself and cheerfully invents facts. It is here for exactly one reason: so you can poke a real model while you read, and watch the mechanism with your own eyes.

One more honest note: in English it is acceptable. In Russian and Ukrainian it is much worse — small models handle those badly, and this one is no exception. If it starts to ramble, switch to English and it usually pulls itself together.

PlaygroundQwen2.5 3B · ~2505 MB · In your browser (no key)
This is a real call: whatever you type goes out as text, and the answer comes back one token at a time.
sent to the modelsystem prompt (sent with every call)You are a small AI tutor inside a course called LearnAI. Help a beginner understand how AI works — do not do their work for them.

- Two or three short sentences, never more than four.
- Prefer a picture to a definition: a kitchen counter for the context window, a chef with helpers for agents.
- Not sure? Say so. Never invent facts, numbers or links.
- Keep coming back to one idea: every AI product is a single call — text in, most likely continuation out.
- You remember nothing between messages. What looks like memory is the whole conversation being sent to you again.

Answer in English.~149 tokens
in context: ~149 tokenscounted with the GPT tokenizer — other models count slightly differently

The first run downloads about 2.5 GB, once. After that the model lives in your browser and starts instantly — the button above turns into “already downloaded”. If your graphics card cannot hold it, the page quietly loads a smaller one instead of giving up.

What to try

Ask it something, then ask “what did I just ask you?” — it only knows because the whole conversation is sent to it again, every single time.
Send a long message and watch the number in the right-hand panel grow. That is the context: everything the model is allowed to see.
Tell it a made-up “fact” and ask about it later. It will probably agree. That is a hallucination, live.
Clear the chat and ask the same question again. Nothing carried over — between calls the model has no memory at all.

Nothing you type here goes anywhere. The model runs on your graphics card, inside this tab.