Neovim with the llama.cpp plugin and heavily quantized qwen2.5-coder with 500 (600?) million parameters. It's almost plug and play although the default ring context limit is way too large if you don't have a GPU.