Qwen3.8-27B memory budget  -  2026-11-03

Pick an exact model file and context, then measure peak memory on the target runtime. Current Unsloth Qwen3.8-27B-UD-Q4_K_XL.gguf at repository revision 4ca720788d1e01f1bff70c033e0d0028fd02e502 is 17,559,178,144 bytes = 16.353 GiB. SHA-256: 3f227079003add2511437e5b1e94812e363385225bf6a9b47b0054a72bc8b01e.

F16 attention cache per token = 16 attention layers x 4 KV heads x 256 dimensions x 2 (key and value) x 2 bytes = 65,536 bytes. Per sequence:
8,192 tokens: 0.5 GiB cache + 16.353 GiB weights = 16.853 GiB.
32,768 tokens: 2 GiB cache + weights = 18.353 GiB.
Two 32,768-token sequences: 4 GiB cache + weights = 20.353 GiB.
262,144 tokens: 16 GiB cache + weights = 32.353 GiB.

These are lower bounds, not measured loaded totals. Add recurrent state, runtime workspaces, OS, other applications, vision and any placement overhead. Cache precision, allocation and offload can differ by runtime. A 16 GiB dedicated GPU is not the same budget as a 16 GB unified-memory Mac sharing memory with the OS; 1 GiB = 1,073,741,824 bytes.
Sources: https://huggingface.co/Qwen/Qwen3.8-27B and https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/raw/4ca720788d1e01f1bff70c033e0d0028fd02e502/Qwen3.8-27B-UD-Q4_K_XL.gguf