744B parameters, streaming off your SSD one word at a time

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

1. 1744B parameters, streaming off your SSD one word at a time0:00
2. 2colibrì is not a model. It is the engine that runs one.0:16
3. 3Why streaming works: only ~5% of the model computes each token0:35
4. 4Where the pieces live0:52
5. 5The per-token path:route → union → place → overlap → learn1:10
6. 6It learns what you use1:42
7. 7So why is it slow?2:06
8. 8The whole engine is one C file2:35
9. 9It was never a size problem. It was a placement problem.2:59