The prompt
7 tokens“We reply within five business …”
The start of a sentence from Larkspur’s guides.
Module 7 · Transformers
What happens inside the model before its next word
A language model writes one token at a time. Each time, every token of the prompt becomes a vector, attention lets each one weigh the tokens before it, and the last one ends up as a score for every token the model knows. Here you open a tiny one up and watch.
Read as
The whole idea in one example
The model scores every next token
What comes next
768 scoredLarkspur mini’s five likeliest, of all 768.
Pick one, add it to the prompt, run again, and a reply grows one token at a time. Larkspur mini is tiny: 51,072 numbers, trained only on Larkspur’s own documents. It runs for real in your browser. It shows how the machinery works; it is no provider’s model.
How it works
Four ideas carry the module
1 One token at a time
The model reads the prompt’s tokens and scores the next one
Vaswani and colleagues introduced the Transformer in 2017, a network built on attention alone, with no recurrence or convolutions.
Source, checked 21 October 2026
2 Attention
Each token weighs the tokens before it
Scaled dot-product attention is softmax(QKᵀ / √d_k) V: each query is scored against every key, and the scores weight the values.
Source, checked 21 October 2026
3 The causal mask
Token 4 looks only back, never ahead
The original Transformer's decoder masks attention so a position cannot look at the positions after it, which keeps generation one token at a time.
Source, checked 21 October 2026
4 Heads and layers
2 layers of 2 heads, each looking its own way
Multi-head attention runs several heads side by side, so the model can attend to different kinds of information at once.
Source, checked 21 October 2026
The story
Larkspur opens up a model
Larkspur Supply Co.’s team opens up “Larkspur mini”, a tiny model trained only on Larkspur’s own documents, and watches it work in the browser. You’re on the team that wants to know why it says what it says, and what a long prompt costs.
Played Tokens? Larkspur mini reads the same tokens as this site’s tokenizer there. RAG chapter 2 showed why you can’t send everything; chapter 6 here shows what a long prompt costs inside.
- 51,072
- numbers in Larkspur mini’s weights
- 2 × 2
- layers × heads
- 32
- tokens in its window
- 44%
- of unseen Larkspur text: the real next token is in its top 5
Your path
The chapters
Chapter 2 · done
Tokens become vectors
How does a token id become something the model can compute with, and how does it know where the token sits?
Chapter 4 · done
Heads and layers
Why does a model have many heads and many layers, and how do their results come together?
Chapter 6 · done
Why long prompts cost more
Why does the work grow faster than the prompt, and what does the KV cache save?
Who it’s for
- Engineers
- Product managers
- Analysts
- No code needed
Engineers, product managers and analysts who use language models and want to see what happens inside one. It builds on Tokens & embeddings; any maths is optional.