Local Gemma4-12B model via llama-server at 127.0.0.1:8081. Free, fast (~40 tok/s), good at math and code. No API key required.