What Models Can My Hardware Run?

How it works, and why it is an estimate

Whether a model runs well comes down to one question: does the whole thing fit in fast memory? The finder answers it in three steps.

The model has to fit in the right kind of memory. That is not the same number as how much RAM or disk you own.

1. Work out your usable model memory. With a dedicated graphics card, that is its VRAM. On a Mac, the CPU and GPU share one pool (unified memory), and macOS lets the GPU use roughly two-thirds to three-quarters of it by default. On a CPU-only PC, it is about 70% of your RAM, because the operating system and your other apps need the rest.
2. Add overhead to each model. A model's download size is only its weights. Running it also needs buffers and a context window, so the finder adds about 1.2GB to each model before comparing.
3. Compare and label. A model that fits with room to spare is "fits comfortably". One that squeezes in is "fits, but tight". On a PC with a GPU, a model that is too big for VRAM but fits when spilled into system RAM is marked as running with partial CPU offload, which works but is much slower.

Why speed is a label, not a number

Off the GPU, generation speed is limited by memory bandwidth, meaning how fast your machine can read the model's weights for every word. A faster memory type or a Max-class Mac moves the label up. I deliberately don't print tokens per second: it varies with the model, the quantization, the context length, and your exact hardware, and a made-up number would be worse than none. Run ollama ps and time a real answer, and trust that over any estimate.

Quick reference by memory size

The best comfortable pick at common sizes, skipping models that would crawl. This table is generated from the same data and rules as the finder (assuming DDR5 for PCs and a base or Pro-class Mac chip), so the two don't disagree.

With a dedicated graphics card

Mac, or a PC with no dedicated GPU

Thinking about buying hardware for a specific model? The GPU pick and the RAM upgrade are the two changes that move you up this table fastest, and the storage guide covers how much disk the models themselves take.

What actually goes wrong

It "fits", but it is painfully slow

Almost always a model that spilled out of your GPU into system RAM. Run ollama ps while it is loaded and read the PROCESSOR column: 100% GPU is what you want, while something like 40%/60% CPU/GPU means part of the model is being computed on the CPU. Pick a smaller model, or see Why Isn't Ollama Using My GPU? if it should be fully on the GPU and isn't.

It works at first, then slows down or crashes in a long chat

The context window grows as the conversation or the documents you feed it get longer, and that extra memory comes on top of what the finder reserved. Models with very large context windows can use far more than the modest allowance assumed here. Shorten the conversation, lower the context length, or step down a size. Ollama Out of Memory walks through the fixes.

You assumed a mixture-of-experts model would need less memory

Models like gpt-oss, Qwen3-Coder, and the "A3B" Qwen releases only use a few experts per word, which makes them fast, but every expert still has to be loaded. The file size is the number that matters for fit. Their "active" size only helps speed.

You have more than one graphics card

Identical cards can often share a single model, but VRAM doesn't simply add up and the finder doesn't model it. Treat the combined total as optimistic. The full PC build guide covers multi-GPU caveats, including why mixing vendors doesn't help.

You only have integrated graphics

Choose "No dedicated GPU". Integrated graphics on a PC share ordinary system RAM, so the model effectively runs on the CPU, and the CPU-only numbers are the honest ones.

Common questions