RAM and VRAM solve different parts of the problem

VRAM is the fast memory attached to a GPU. System RAM is the working memory used by Windows, applications and CPU-side model execution. A local AI setup can have plenty of VRAM and still run poorly if system RAM is nearly full. The reverse is also true: a machine with a lot of RAM can often load a model through CPU execution or partial GPU offload, but it may be much slower than a setup that keeps more of the model on the GPU.

For that reason, RAM should be treated as operating headroom rather than as a direct measure of AI quality.

What consumes system RAM?

The obvious answer is the model, but that is only part of it. Windows itself needs memory. So does the local runtime, the user interface, browser tabs, coding tools, document parsers, retrieval indexes and any other application open at the same time. Long contexts and agent workflows can add further memory pressure.

If the model does not fit fully in GPU VRAM, many runtimes can keep some model layers or other state in system RAM and execute part of the workload on the CPU. That makes RAM especially important on machines with integrated graphics, smaller GPUs or workflows that deliberately use CPU offload.

Practical RAM bands for Windows local AI

These are conservative starting points, not compatibility guarantees:

  • 8 GB: very constrained. Windows and normal applications already consume a meaningful part of the available memory. Small experiments may work, but there is little room for offload, long context or multitasking.
  • 16 GB: a realistic entry point for smaller local models and lighter workloads, particularly when the GPU has enough VRAM to carry most of the model.
  • 32 GB: a comfortable general-purpose range for many local-AI users. It leaves more room for CPU offload, coding tools, document work and multiple applications.
  • 64 GB: useful for larger quantised models, heavier CPU-assisted workloads, long-running development environments and users who want substantial headroom.
  • 96–128 GB and beyond: specialist territory for larger CPU-resident models, heavier experimentation, multiple runtimes or workstation workloads that already need large memory capacity outside AI.

The correct band depends on the model and workload. A 16 GB machine with a capable 16 GB GPU can behave very differently from a 32 GB machine with no discrete GPU.

Do not size RAM to the model file alone

A downloaded model file is not a complete memory budget. Runtime allocations, context state and the rest of the operating system still need space. Memory-mapped model loading can also make task-manager numbers look less intuitive than a simple “file size equals RAM use” rule suggests.

Leave headroom. A machine that technically loads the model while Windows is paging heavily to storage is not a good interactive local-AI experience.

The page file is not a RAM upgrade

Windows virtual memory is valuable as a safety mechanism, but storage is far slower than physical RAM. A large page file can prevent an immediate out-of-memory failure; it cannot turn an undersized memory configuration into a fast model workstation. If an interactive model continuously pushes the machine into heavy paging, the practical fix is usually a smaller model, more aggressive quantisation, more physical memory or better GPU fit.

Integrated graphics can share system memory

On systems without dedicated VRAM, the graphics processor may use shared system memory. That means the same RAM pool has to serve Windows, applications and graphics/AI workloads. Capacity therefore matters even more, although runtime and driver support still determine whether the integrated GPU can accelerate a particular model effectively.

Context and agents increase the need for headroom

A short chat is not the same workload as a coding agent reading a repository, a document assistant with a large context or a tool-using agent running alongside a browser. The larger the working set around the model becomes, the more useful spare RAM is.

For workstation planning, think about the whole workflow: model plus runtime plus context plus tools. HardwareRadar focuses on the actual machine, while LLMRadar is designed to connect hardware constraints with model and task selection.

RAM or VRAM: which should you upgrade first?

If system RAM is already close to full during normal work, more RAM can improve the entire machine. If the system has comfortable RAM headroom but most model layers are falling back from GPU to CPU, more VRAM can make a much larger difference to inference speed. The answer depends on where the bottleneck actually is.

Use the Local AI Readiness Estimate for a first-pass fit check, then read How Much VRAM Do You Need for Local AI? and Which Local LLM Can My PC Run? together.

Bottom line

For most Windows users, 16 GB is the practical floor for serious experimentation and 32 GB is a much more comfortable general-purpose target. Larger capacities become increasingly useful when models spill into system RAM, contexts grow or the PC is doing real work alongside the model. Treat those numbers as planning bands, not promises.