After listening to various perspectives across numerous threads, I’ve encountered a wide range of experimental approaches. I invite you to share your setups here as well, so we can try to identify the absolute best configuration. The best coding setup I’ve seen so far is Qwen 3.5 27B 8-bit + llama.cpp + async KV cache (K=Q8, V=Turboquant—I learned about this from an Alex Zistand video).