A local LLM needs more than the download size. For Qwen2.5-7B-Instruct, the BF16 safetensors selected for that checkpoint total 15,231,233,024 bytes (about 14.18 GiB). That figure comes from model.safetensors.index.json metadata on Hugging Face at revision a09a35458c702b33eeacc393d103063234e8bc28. It is the weight file payload, not peak GPU memory.
On top of weights, the KV cache for this architecture (28 layers, 4 KV heads, head dim 128) at 8,192 cached tokens, 1 sequence, FP16 KV is:
2 × 28 × 4 × 128 × 8192 × 1 × 2 = 469,762,048 bytes (0.44 GiB of raw KV).
Runtime extras (workspace and engine reserve) are scenario ranges, not a hidden 1.2× factor. With calculator 0.1.0, dataset 2026-09-19.1, llama.cpp, typical extras, and no extra safety margin, the total estimate is a range above the 14.18 GiB weights — not a single fake-precise number.
The same model in official GGUF Q4_K_M (two shards only, 4,683,073,632 bytes) drops the weight term to about 4.36 GiB. KV does not automatically become 4-bit when weights are 4-bit.
Conditions for the numbers on this page: Qwen2.5-7B-Instruct, one sequence, 8,192 cached tokens, FP16 KV unless stated, no tensor parallel. Open the calculator with that model to recompute.