#553 · +72 · 1d ·GitHub
md). It is a fork of TotK Recomp by Valiantsin Dzerakh.
Global
IMPORTANT] > This project was made almost exclusively with AI.
\janas-chat. 8 times as fast with the dense model (Qwen3-4B 346 tok/s against 193), whose whole blocks go to the GPU, through Intel's own Windows driver there.
#1789 · +15 · 1d ·GitHub
24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.
JSON is everywhere on the Internet. Servers spend a lot of time parsing it. We need a fresh approach.