Yep, Gated DeltaNet in Qwen3.6 requires much less VRAM for the KV cache than previous generations. Plus the KV cache is 8-bit.

is it in llama.cpp?