drumih/turbo-fieldfare
Runs Google's Gemma 4 26B model on any Apple Silicon Mac in about 2 GB of memory, including the 8 GB machines, instead of loading all 14.3 GB. It keeps the shared core and the cache resident and streams only the experts each token actually needs off the SSD. The runtime, installer, command line tool and Mac app are all written in Swift and Metal, built for this one model rather than wrapping MLX or llama.cpp.
Stars 3,038 to 3,848, up 26.7%, with forks up 38.8% and open issues up 66.7%. Forks and issues both outpacing stars means people are running the code and hitting real edges. HackerNews has gone quiet at 1.5%.