Ollama does this, as does llama-server from llama.cpp