You can run Qwen3 locally today if you want to. It can write whole files if you want (although not with <1 second latency like a sub 1GB model will which is what you want for interactive in-editor completions.)