Step 1 of 5 · Follow one token
Last time I analyzed Qwen mainly from its outputs. After our discussion, I wanted to understand what is happening inside the model. Here is one sentence, followed token by token.
Recorded Qwen measurement· real tokenizerClick a token to follow it.
“␣notebook” → id 37166 → one row of the embedding table → a vector of 4,096 numbers in Qwen.
its starting vector (8 numbers) Toy calculation
From here the vector passes through 36 blocks, a final RMSNorm and the output layer, which gives a score for each of 151,936 vocabulary rows.
Keys: ← → move between steps, N notes, D deeper sections. Nothing advances automatically.