Tome Sweet Home

Explore AI

What can your machine run?

Find the hardware you already own. Each tier says what fits, what does not, and what it will feel like.

MEMORY A MODEL NEEDS: Qwen3 1.7B 1.5 GB, Qwen3 4B 3 GB, Qwen3 8B — the usual answer 5.5 GB, Qwen3 32B 20 GB, 70B class 40 GB

The single most common question about local AI is "can my computer run this?" — and the answer is arithmetic, not opinion.

The arithmetic

A model's memory footprint is roughly:

weights = parameters × bytes per weight   (at Q4, about 0.55 bytes per parameter) > > plus the KV cache — the model's working memory of your conversation, which grows with context length and can add several gigabytes > > plus overhead for the runtime itself

So an 8B model at Q4 is about 5.5 GB of weights. Add context and overhead and you want 8 GB free to be comfortable. A 32B model is about 20 GB, so you want 24 GB.

That is the whole trick. Everything below follows from it.

Pick your machine

The tiers below are shortcuts. Each one lists what genuinely fits — not what fits if you close every other application and accept a two-minute wait.

A warning about "it runs"

Plenty of guides will tell you a model "runs" on your hardware. What they usually mean is that it loads and produces tokens. That is not the same as being usable:

  • Below about 5 tokens per second, conversation feels broken. You are reading slower than you think.
  • Quantised below Q4, quality drops in ways that benchmarks do not show.
  • Offloading to system RAM makes a model fit, at a speed that usually makes it pointless.

The tiers here set a higher bar: they list what you can use without resenting it.

If nothing here fits

That is a real answer, and covered honestly in the last tier. Renting a GPU by the hour, or using a hosted model, is not a failure — for most people it is the correct engineering decision.

Mac with 8 GB Workable, but only just — and not as a daily driver for anything serious. 3 models fit Mac with 16 GB The real starting point for local AI. 3 models fit Mac with 32 GB The comfortable tier. This is where 30B-class models become practical. 4 models fit Mac with 64 GB or more You can run almost anything that is open, at some speed. 4 models fit NVIDIA GPU with 8 GB VRAM Fine for small models, tight for 8B. 3 models fit NVIDIA GPU with 12 GB VRAM The sensible budget tier for local AI. 3 models fit NVIDIA GPU with 16 GB VRAM Opens up Mixture-of-Experts models, which change the economics. 3 models fit NVIDIA GPU with 24 GB VRAM The enthusiast tier. 30B-class models at quality quantisations. 4 models fit 48 GB VRAM or more Large open models, no compromises. 3 models fit No GPU at all Possible, and slower than you would like — but not useless. 3 models fit I have nothing suitable, and do not want to buy Use a hosted model, or rent a GPU by the hour. 0 models fit

What changed, and when →