VRAM CALCULATOR

Before you download 40 GB of weights, do the napkin math properly. Pick from 50+ open-weight models, choose a quantization, then drag the sliders — context length, KV-cache precision, CPU offload — and watch exactly how much VRAM and system RAM the run would need, broken down into weights, KV cache, and runtime overhead. If the verdict says it doesn’t fit, it doesn’t fit — optimism is not a memory technology.

Model

Quantization

Context length

1Kmax

KV cache precision

CPU offload — layers in system RAM

0% · all GPU100% · all CPU

VRAM needed

–

System RAM needed

–

Footprint vs budget

Weights KV cache Overhead

Estimates, not gospel. Weights = total params × effective bits-per-weight / 8 (GGUF quants carry metadata, so Q4_K_M ≈ 4.85 bpw, not 4.0). KV cache = 2 × layers × KV heads × head dim × context × bytes-per-element. Overhead = 0.35 GB fixed + 2% of weights for activations and compute buffers. Real footprints move with runtime, batch size, and build — treat these numbers as a solid planning floor, then leave headroom.