Georgi Gerganov, the creator of llama.cpp, runs Qwen3-27B locally for daily coding work. He uses it on an M2 Ultra and an RTX 5090, handling routine maintenance tasks at ggml-org. His commit history is public proof.
His setup is deliberately minimal: the pi agent stripped to its core with the offline flag and a short system prompt tuned to his personal style. No elaborate scaffolding. The person who built the infrastructure everyone else relies on keeps his own stack lean.
The original Hacker News comment is worth reading because Gerganov is candid about the ceiling: PR review load limits how much he can actually use the tool. That constraint, from that source, tells you more about the current state of local AI coding assistants than most benchmarks will.
[READ ORIGINAL →]