Training-free mixed-precision Qwen3.5-4B for 12 GB GPUs. With an 8-bit KV cache: 96% of BF16's score in 30% of the memory, 2.7x faster.