A user reports running the nvidia/Qwen3.6-27B-NVFP4 model as a daily driver for agentic coding tasks on an NVIDIA DGX Spark system. The configuration utilizes compressed-tensors quantization and fp8 kv-cache dtype to operate within the hardware's memory constraints. Key settings include a max model length of 131072, speculative decoding via MTP with 3 tokens, and auto tool choice enabled.