Qwen 3.6 27B is a BEAST

I have a 5090 Laptop from work, 24GB VRAM. I have been testing every model that comes out, and I can confidently say I’ll be cancelling my cloud subscriptions. All my tool call and data science benchmarks that prove a model is reliably good for my use case, passed. It might not be the case for other professions, but for pyspark/python and data transformation debugging it’s basically perfect. Using llama.cpp, q4_k_m at q4_0, still looking at options for optimising. submitted by /u