Lots of iterating on my local LLM setup. Really happy with the current state. Qwen3.6-27B-IQ4_NL.gguf from Unsloth, 128k context window, vision capability, 50 tokens per second, surprisingly good quality output in agentic dev. All in 24gb of VRAM. The MoE variant still has its place (120 t/s!) but this one is the default now.