Which AI Models Fit Your Laptop? One Free Command Answers

“Can my laptop run AI locally?” is the most common question I get from students — and the usual answer is a shrug and a 40-minute download that turns out too slow to use. There is a one-command way to know before you download anything.

The tool: llmfit

llmfit detects your machine’s RAM, CPU and GPU, then scores every open model in its database for fit, speed and quality — right-sizing models to your hardware. It runs locally, needs no API key, and works on Apple Silicon, NVIDIA and AMD.

Install (one line)

brew install llmfit

(Or the cross-platform installer on the repo.)

See what your machine can run

First, what llmfit sees:

llmfit system

It prints your CPU, total and available RAM, memory bandwidth, and GPU. Then the ranking:

llmfit

You get a table of models, each marked 🟢 Perfect / 🟡 Good / red (won’t fit), with estimated tokens/sec, memory use and quantization. Green means download it today. Red means don’t waste the afternoon.

Checking a specific laptop (e.g. 8 GB)

llmfit lets you override the detected hardware, so you can plan for a machine you don’t own yet:

llmfit --ram 8G

On an 8 GB laptop, the top fits are small reasoning models — DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-0528-Qwen3-8B, and Qwen3-4B — all comfortably green. Filter by name (press / in the interactive view and type deepseek) and llmfit shows exactly how many of that family fit.

My 3 starter picks by RAM

• 8 GB: DeepSeek-R1-Distill-Qwen-7B (reasoning that actually fits) or Qwen3-4B for speed.
• 16 GB: an 8B model at higher quantization — DeepSeek-R1-0528-Qwen3-8B is a strong daily driver.
• ≥32 GB / good GPU: step up to 14B+ models; run llmfit and take the top 🟢 rows.

Pair any of these with Ollama (ollama run <model>) and you have a private, offline AI on your own machine.

The honest limits

• Fit ≠ fast enough. llmfit estimates tokens/sec; a model can “fit” and still feel slow. Check the tok/s column, not just the green dot.
• It’s an estimate. Real speed depends on quantization, context length and background load.
• It ranks open models, not closed APIs — this is about what you can run yourself, free.

The next time someone asks “will this run on my laptop?”, you don’t guess — you run one command and know.