Topic: linear-attention

FareedKhan-dev/

kimi-k3-in-c

A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU.

serphen/

microkimi

Run and train your own miniature Kimi K3: a zero-dependency Rust inference engine, reimplemented from scratch - and nanokimi, a small model trained from scratch on CPU by the included training pipeline. Talks straight to the Metal API for GPU on macOS - CPU mode stays the fastest at small model sizes.