Supporting fit tool for 313 models. Default Q4_K_M · 4k · batch 1. Estimate, not a bench.
Dense 7B–70B totals sit ~5–8% above the Q4_K_M GGUF file — planning number, not peak runtime VRAM.
We may earn a commission from cloud and hardware partners at no extra cost to you.
Estimate · Q4_K_M · 4k context · batch 1. For dense 7B–70B this sits ~5–8% above the Q4_K_M GGUF file. Not peak runtime VRAM and not a measured bench.
MoE / PLE load uses total (storage) parameters. Active or effective counts are architecture metadata, not the VRAM formula.