Home › Can my computer run it?

Can my computer run it?

Pick your graphics card or Mac and see which popular open-weight AI models fit in its memory, at three common quantizations. Sizes come from each model's files on Hugging Face, and the few without a file are worked out from the parameter count. These are estimates: real use depends on the app, the context length and your settings.

Pick your hardware to see what fits.

Nothing in our list fits at Q4. Try a smaller amount of context, or turn off "Only show models that fit" to see what is close.

How we estimate

• Model sizes: the exact size of each model's Q4_K_M, Q8_0 and FP16/BF16 GGUF files, read from the Hugging Face API. Where no FP16 GGUF is published we show parameters × 2 bytes and mark it "computed".
• Parameters: the parameter count Hugging Face lists for the original model (its safetensors total).
• Fits: file size plus about 1.5 GB is at or below your usable memory. Tight: it fits, but within 10% of the limit. Won't fit: more than your memory; on a graphics card it may still run with part of the model on the CPU, but much more slowly.
• Macs: we count about 75% of unified memory as usable, because macOS keeps the rest for itself by default. No graphics card: we also count about 75% of system RAM, leaving room for the OS and apps (our assumption); it runs, but slowly.
• Q4 is the Q4_K_M file; where a repo only publishes an Unsloth Dynamic version we use that and name it (for example UD-Q4_K_XL). When a model only fits as a much smaller build (1- to 3-bit), we say so: those builds lose noticeably more quality.
• Mixture-of-experts models (marked MoE) still need all their weights in memory, even though only part of them is used for each word.
• Longer context (for example 32K or 128K tokens) needs more memory than the 8K we assume. Speed is not estimated here.

Model data checked on Hugging Face on Oct 16, 2026. 31 models. Sizes are estimates, not a guarantee. See also How we verify.