How it works

How the kiwi models are made

Six short chapters, each built around one moving figure. Every number in them was measured by the project.

  1. Inside the model Each token climbs every layer 28 Layers in Flyingkiwi 0.6B
  2. Training A teacher’s answers, cut short 35B → 0.6B Teacher and student
  3. Making it small 256 levels or 16 256 / 16 Levels at 8-bit and 4-bit
  4. In your browser Checked parts, sixteen lanes 17 Parts of up to 32 MiB
  5. The library 49,690 pages, one key 49,690 Wikipedia summaries, offline
  6. Answer formats Locked keys, kept layouts 6 Answer formats