Free tool
Will this model fit on my GPU?
Pick a card, a model, a quantisation and a context length. You get weights, KV cache and runtime overhead separately, plus an honest fits-or-spills verdict. The maths is calibrated against real measurements taken on an RTX 3090, not vendor marketing.
| Model weights | — |
| KV cache | — |
| Runtime overhead | — |
| Total | — |
| Context | KV cache | Total | Verdict |
|---|
Calibrated, not guessed. Two real runs on an RTX 3090 with qwen3.8:27b,
q8_0 KV cache and flash attention: context 32,768 used 20,373 MiB, and context
98,304 used 23,101 MiB. Weights come from the Ollama API (17.7 GB, Q4_K_M).
Fitting those two points gives roughly 2,129 MiB of fixed overhead plus
42.6 KB of KV cache per token.
It predicts its own failure case. The same fit puts context 131,072 within
about 0.1 GB of the card's 24 GiB — and that is exactly the setting where the
real run spilled onto the CPU.
What is estimated. Other models scale KV cost linearly from that anchor by
parameter count, so treat non-qwen3 numbers as a close estimate rather than a
measurement. Quantisation ratios are the standard llama.cpp bits-per-parameter
figures. Leave a little headroom: your desktop needs VRAM too.