Skip to main content

Methodology explainer

Local AI and VRAM: why UpgradeOrSave does not yet say which model fits

Whether a local model runs well depends on model, quantization, context and backend. Without verified runtime profiles, we make no fit claims.

By UpgradeOrSave EditorialEditorial update Published

Editorial dates describe the text. They do not date the measurements. Measurement acquisition dates are unavailable in this view.

One of the most common upgrade questions today is whether a graphics card can run a particular local language model. It sounds like a lookup: compare the model's size with the card's memory and you are done. In practice the answer depends on several choices made when the model is run, and getting them wrong produces confident advice that does not survive contact with a real setup.

What actually decides whether a model fits

  • The exact model and checkpoint, because files with similar names can differ.
  • The quantization, which changes how much memory the weights need and how quality behaves.
  • The context length, because the working memory for the conversation grows with it.
  • The runtime and backend, which differ in memory overhead and in which GPUs they support well.
  • Everything else on the card, including the desktop and other applications.

Change any one of these and the memory needed, and the speed you get, can change substantially. A fit claim without them is a guess.

A file size is not a requirement

We do hold reviewed metadata for some official model checkpoints, including their exact file sizes. That proves which file we are talking about. It does not prove how much video memory a given runtime needs at a given context length, or whether a specific GPU and backend can run it at a usable speed. Treating checkpoint bytes as a VRAM requirement would turn a precise fact into an imprecise promise.

What we need before making fit claims

A fit claim needs a verified execution profile: a source-backed record of a model, quantization, context, runtime and backend, with the memory it actually used and the speed it achieved on identified hardware. Until such profiles are reviewed and admitted, our Local AI pages state the gap plainly and show model identity facts only.

What you can do today

Use the Local AI section to see which reviewed model artifacts exist and what is known about them. For a specific setup, the runtime you plan to use is the best source of its own memory behaviour. When verified profiles are admitted, upgrade checks will be able to report capability changes alongside performance, and this explainer will be updated.

What this means for an upgrade

  • Do not choose a card for local AI from a VRAM rule of thumb presented as a measurement.
  • Check memory needs with the exact model, quantization, context and runtime you plan to use.

Limitations and uncertainty

  • No verified local AI execution profile is currently admitted, so no model-fit or VRAM threshold is stated anywhere on UpgradeOrSave.

Evidence and methodology

Source-rights checks and active research holds apply before evidence reaches this page. An unavailable result is omitted rather than estimated.

No commission-tracked links appear in this article. Any future affiliate links require separate activation and a clear disclosure near the link. Read our disclosure.