Interactive · Local AI

Will It Run? — Local AI on Your PC

A real AI model can run entirely on your own machine — no cloud, no subscription. The only question is which size fits. Pick your hardware below and see which models run smoothly, which are usable but slow, and which won't fit.

GPU VRAM (the fast path)

System RAM

…

Memory budget

How to read this.

Smooth — the model fits entirely in your GPU VRAM (or your Mac's unified memory) with headroom, so it runs on the GPU — fast.
Usable — too big for VRAM but it fits in system RAM, so it runs on the CPU or partially offloaded — it works, just slower.
Won't fit — not enough memory to load it.

Sizes assume Q4_K_M quantization (the common "good enough, 4× smaller" format), at a modest context length. Real numbers vary with context size and the exact build — treat this as a planning estimate, not a guarantee.

Plain-English glossary

VRAM: Your graphics card's own fast memory. A model that fits here runs on the GPU — the quick path.
Unified memory: On Apple Silicon, the CPU and GPU share one memory pool, so "RAM" and "VRAM" are the same thing — which is why Macs punch above their weight for big models.
Parameters (B): A model's size in billions of weights. More parameters usually means smarter, but needs more memory.
Quantization (Q4): Compressing those weights to ~4 bits each instead of 16 — roughly 4× smaller, with a small quality trade-off. It's what makes local AI fit on normal hardware.
Tokens / sec: How fast the model writes. On the GPU it feels instant; on the CPU it can be a slow crawl.
Context window: How much text the model can "hold in mind" at once. Bigger context eats more memory on top of the model itself.