Teapot

I'm a teapot.

What this is

Teapot is a deliberately tiny language model, an homage to HTTP 418 "I'm a teapot." It has about 128,000 parameters and was trained on a single riddle paragraph plus a small set of questions derived from that riddle's own lines. It has no general knowledge and cannot answer questions outside its training data. That is the point.

The project exists to showcase serving infrastructure, not model intelligence. Teapot runs through the same production serving stack as every other model in the lineup: continuous batching, an efficient KV cache, streaming, and a standard OpenAI compatible chat API. The model is intentionally trivial so the stack around it is the actual subject.

The riddle

I'm a teapot, short and stout, asked for something I won't pour out. A message arrives with a purpose in doubt, but one little number will give me away outright. I simmer, I signal, I stubbornly stay, refusing the task in a peculiar way. Not broken, not lost, just tied in a knot, what am I, brewing inside the pot?

Usage

Teapot answers a small fixed set of questions related to the riddle above (for example "are you stout?" or "will you pour?") and can recite the full riddle on request. Anything outside that scope will produce an unpredictable or nonsensical answer, since the model was never trained on anything else.

from mlx_lm import load, generate

model, tokenizer = load("terion-mlx/teapot")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "tell me a riddle"}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt))

Architecture

A tiny Llama style transformer, natively supported by mlx_lm with no custom model code: 3 layers, hidden size 64, 4 attention heads with 2 key value heads, a 258 token byte level vocabulary built from scratch on the training text. Total size on disk is about 500KB.