Local LLMs for code completion: worth the setup?

I've been tinkering with local models like Qwen2.5-Coder and DeepSeek-Coder for IDE autocomplete, mainly to avoid sending proprietary code to cloud APIs. The privacy angle is great, but I'm hitting a wall with context size and latency on my RTX 4060. Even with llama.cpp and quantized weights, it feels slower than Copilot, and the suggestions are often more generic. For those of you running local coding assistants daily, what's your setup? Are you using agentic workflows or just tab-completion? And have you noticed a real productivity boost, or is it mostly the privacy win keeping you on it? Curious if the trade-off is actually worth it for larger codebases.