Free tool · 100% on-device
Will this model fit on my GPU?
Set your card, model, quant and context length below. It splits the total into weights, KV cache and runtime overhead, then gives you a straight answer: fits, or spills to CPU. The formula comes from two timed runs on a real RTX 3090, not a spec sheet.
Your setup
Where the memory goes
| Model weights | — |
| KV cache | — |
| Runtime overhead | — |
| Total | — |
Context ladder
| Context | KV cache | Total | Verdict |
|---|
Where these numbers come from: two timed runs on an RTX 3090, running qwen3.8:27b with q8_0 KV cache and flash attention on. At 32,768 tokens of context, VRAM use hit 20,373 MiB; at 98,304, it hit 23,101 MiB. Weight size (17.7 GB, Q4_K_M) is pulled straight from the Ollama API. Fit a line through those two points and you get about 2,129 MiB of fixed overhead plus 42.6 KB of KV cache per token — and that same line puts a 131,072-token run within 0.1 GB of the card's 24 GiB ceiling, which is exactly where the real run spilled onto the CPU.
Everything outside the qwen3 family is a scaled estimate, not a second measurement: KV cost scales linearly from that anchor by parameter count, and the quant ratios are the standard llama.cpp bits-per-parameter figures. Treat the numbers as a strong estimate, and leave yourself some headroom — your desktop needs a slice of that VRAM too.