The finder covers MacBook Air, MacBook Pro, Mac mini, iMac and Mac Studio with Apple M1 to M5 chips, and Windows PCs with NVIDIA RTX or AMD Radeon graphics cards, or none at all. It suggests open models you can run locally for free with Ollama, from small ones for 8 GB laptops to large LLMs for 128 GB workstations.
Best local AI model for popular computers
The best all-round pick for some of the most common setups, with an estimate of how fast it answers.
| Computer | Best model | Speed (estimate) |
|---|---|---|
| MacBook Air M1, 8 GB | Qwen 3.5 4B | about 10 words a second |
| MacBook Air M2, 16 GB | Qwen 3.5 9B | about 8 words a second |
| MacBook Air M3, 16 GB | Qwen 3.5 9B | about 8 words a second |
| MacBook Air M4, 16 GB | Qwen 3.5 9B | about 10 words a second |
| MacBook Air M5, 24 GB | GPT-OSS 20B | about 25 words a second |
| Mac mini M4, 16 GB | Qwen 3.5 9B | about 10 words a second |
| MacBook Pro M4 Pro, 24 GB | GPT-OSS 20B | about 40 words a second |
| MacBook Pro M4 Pro, 48 GB | Qwen 3.5 27B | about 8 words a second |
| MacBook Pro M5 Max, 128 GB | GPT-OSS 120B | about 65 words a second |
| PC, no graphics card, 16 GB | Qwen 3.5 9B | about 5 words a second |
| PC, RTX 3060 12 GB, 32 GB | GPT-OSS 20B | about 25 words a second |
| PC, RTX 4060 8 GB, 16 GB | GPT-OSS 20B | about 15 words a second |
| PC, RTX 4070 12 GB, 32 GB | GPT-OSS 20B | about 25 words a second |
| PC, RTX 4090 24 GB, 64 GB | Qwen 3.5 27B | about 30 words a second |
How the finder decides
- Will it fit? A model needs about its download size in memory, plus room for the conversation and your other apps. On a Mac, we keep the model under 60% of memory, the same rule PetekMind Hub uses. On a PC, the fastest place is the graphics card’s own memory; a model that doesn’t fit there can still run partly in system memory, more slowly.
- How fast will it be? To write each word, the computer reads through the model once, so speed comes down to how fast your chip or graphics card can move data through memory. That’s why an M4 Pro runs the same model about twice as fast as a plain M4. Some models, like GPT-OSS, only read part of themselves for each word, so they’re faster than their size suggests.
- Which one is best? For everyday use, we pick the biggest strong all-rounder that fits and still answers at least as fast as you read. For photos, maths or speed, we pick the best model for that job instead.
The commands shown are for Ollama, but the same models also run in LM Studio and other apps, including Apple’s MLX versions on a Mac.
More detail, including how to check your memory, is in our guide to the best local AI model for your computer.
Questions
Which local AI model is best for a MacBook Air?
It depends on the memory. With 8 GB, Qwen 3.5 4B is the best all-rounder. With 16 GB, Qwen 3.5 9B gives clearly better answers. With 24 GB, GPT-OSS 20B fits, and with 32 GB Qwen 3.5 27B. Newer chips run the same model faster.
Can a Mac with 8 GB of memory run local AI?
Yes, small models. Qwen 3.5 4B runs well on an 8 GB M1, at about 10 words a second. It’s good for short questions, emails and summaries. For better answers, 16 GB or more makes a big difference.
Is 16 GB of memory enough for local AI?
Yes. 16 GB is the sweet spot for an everyday assistant: on a Mac it runs Qwen 3.5 9B, which handles writing, rewriting and summaries well. 32 GB or more opens up the top open models.
What local AI model can an RTX 4060 run?
The RTX 4060 has 8 GB of graphics memory. Smaller models such as Qwen 3.5 9B run fully on the card and are very fast. With 16 GB or more of system memory, GPT-OSS 20B also runs, split between the card and system memory.
How accurate are the speed estimates?
They’re rough estimates, best for comparing options rather than as exact numbers. Speed depends mostly on how fast your chip or graphics card moves data through memory, which is published for each chip and card. Long conversations, other busy apps and laptop power settings slow it down.
Does the model finder send my information anywhere?
The finder runs in your browser. We count which options people pick, without names or email addresses, so we know which computers to cover next.