LFM2-24B-A2B runs at 112 tok/s on CPU with only 2.3B active params. Not a transformer. GGUF files from 13.5GB, Ollama and llama.cpp setup, and where it beats Qwen.