Free · runs in your browser · no sign-up

Learn how LLMs work by running one in your browser

Chat with a small language model on your own GPU, or plug in any OpenAI-compatible API. Then look inside: the system prompt, the tokens, the context window, every tool call and how fast it thinks.

  • WebGPU, on your machine
  • 5 models ≤ 2B
  • Any OpenAI-compatible API
agent loop · qwen3-1.7b
  1. youWhat is 23 × 19?
  2. model<tool_call>{"name":"calculator","arguments":{"expression":"23*19"}}</tool_call>
  3. toolcalculator → 4374 ms · sandbox
  4. answer23 × 19 = 437
Context 486 / 4,096
Decode 38 tok/s
The model decides to call a tool, the tool runs in a sandbox, the model answers. You see every step.

Inside the model

What you'll see that chat apps hide

How it works

Three steps, no setup

  1. Pick a model

    One of 5 small models (0.6B–1.7B parameters) downloaded once into your browser, or any OpenAI-compatible API such as OpenAI or Ollama.

  2. Ask something

    Change the system prompt, switch tools on or off, or write your own tool with a JSON Schema and a few lines of JavaScript.

  3. Watch what happens

    Follow every step of the loop, and see the context window and tokens per second update as the model generates.

Learning path · 8 lessons · 27 min total

Learn the concepts

All lessons
  1. 01

    What is a token?

    Models read pieces of words mapped to numbers, not words.
    3 min
  2. 02

    Temperature, top-p and top-k

    How the model picks one next token out of thousands of options.
    4 min
  3. 03

    The context window

    The model's working memory, and what happens when it fills up.
    3 min
  4. 04

    Managing the context window

    Dropping, trimming and summarizing: how chat apps make room, and what each one forgets.
    4 min
  5. 05

    The system prompt

    The hidden instructions sent before every conversation.
    3 min
  6. 06

    How function calling works

    A model can only write text. Here is how that text becomes a tool call.
    4 min
  7. 07

    Prefill vs decode

    The two speeds of every LLM: reading your prompt and writing the answer.
    3 min
  8. 08

    Run an LLM in your browser

    WebGPU, small models, and what to expect from a 1B model on your laptop.
    3 min

FAQ

Questions

Is it free?

Yes. It runs in your browser and the code is on GitHub. Local models cost nothing to run. If you connect a paid API, you pay that provider directly.

What do I need to run a model locally?

A browser with WebGPU (recent Chrome or Edge) and roughly 1–2 GB of free GPU memory. The model downloads once and is cached.

Is my data sent anywhere?

With a local model, nothing leaves your computer. With an API, your messages go only to the API address you enter, and your key is never saved: you enter it each visit.

Which models can I use?

Locally: Qwen3 0.6B, Llama 3.2 1B, Qwen2.5 1.5B, SmolLM2 1.7B and Qwen3 1.7B. Remotely: any model behind an OpenAI-compatible /chat/completions endpoint.

See what an LLM is actually doing

Load a model, ask a question, and watch every step. It takes about a minute.