sirfyyn

sirfyyn (Max Bitzer). A private home lab, and FLE (Fast Local Experts): a llama.cpp fork that runs an 89 GB Mixture-of-Experts model on one 32 GB GPU at about 100 tokens/s.

New: "Making Rhea Fast", eight short films on how we made it fast, how we checked that it still answers as well, and why we release all of it.

Code: https://github.com/sirfyyn/fle.cpp
Models: https://huggingface.co/sirfyyn/Qwen3.8-Flash-Next-FLE-GGUF
Website: https://sirfyyn.de

A private project, written with Claude. Hints and findings are welcome; changes may take a while. The older gaming videos stay up.