kimi-k3-in-c

Inference engine written in portable C99 runs trillion-parameter Kimi K3 models on a single CPU with about 8GB RAM, using disk streaming and no BLAS, framework, or GPU dependencies.