GPU MEMORY

How much VRAM do you need for local AI?

Enough to hold model weights, attention cache and runtime headroom for the workload you actually want to run. There is no useful single VRAM number for “AI”.

8 GB

Useful for smaller quantised models and shorter contexts.

12 to 16 GB

Gives more room for medium-sized quantised models, longer contexts and fewer borderline memory situations.

24 GB and above

Opens more larger-model options and longer contexts, but extra capacity is only useful if your target workload needs it.

Model size is only part of the calculation

Quantisation changes weight size. Context length changes cache use. Runtime overhead and other applications consume memory too.

System RAM still matters

Even if a model fits in VRAM, loading can need substantial system RAM. Our calculator reports both targets.

Compare capacities under the same workload

Our 8, 12, 16 or 24 GB GPU memory guide uses a worked example, and the homepage upgrade comparison shows what additional models a larger configuration unlocks.

Next: try the upgrade comparison.