Qwen2.5 VL 7B
Show it a screenshot and it will tell you what is going on. Fits in 8 GB.
Step one: find your memory
On a Mac
Open the Apple menu and pick About This Mac. Look for Memory. Macs with Apple chips share this memory with the graphics, which works well for AI.
On Windows
Open Settings, then System, then About. Look for Installed RAM. If you have a graphics card, check its VRAM in Task Manager under Performance, then GPU.
LM Studio says 16 GB of RAM is a good start. See their system needs.
Step two: pick your size
Each model needs room for its file plus some working space. We add it up with a simple formula. It is a rough guide, so leave some spare room. See the formula.
8 GB of memory
20 models fit. Here are the biggest ones that fit, since bigger usually means better answers.
Qwen3 8B
Thinks step by step on a normal laptop. Fits in 8 GB.
DeepSeek R1 0528 8B
A newer small thinker. MIT license, so legal can relax.
Granite 3.3 8B
IBM's business model. It wears a tie to the chat.
Gemma 4 E4B
Small and efficient. Your laptop will barely notice it is there.
Hermes 3 8B
Built on Llama 3.1, so Meta's rules come along for the ride.
16 GB of memory
28 models fit. Here are the biggest ones that fit, since bigger usually means better answers.
Qwen3 14B
The 8B's older sibling. Smarter, and wants about 13 GB.
DeepSeek R1 Distill 14B
Thinks out loud before it answers. Get comfy.
Phi-4
Good at math, short memory. Keep your chats short and sweet.
Mistral Nemo 12B
Speaks many languages. Fits a 16 GB machine.
Gemma 3 12B
A good writer that can also look at pictures. Wants about 11 GB.
Gemma 4 12B
Google's open model with an Apache license. Fits a 16 GB machine.
32 GB of memory
40 models fit. Here are the biggest ones that fit, since bigger usually means better answers.
Qwen3 32B
A careful thinker. Wants about 26 GB and a little patience.
Olmo 3.1 32B
Largest fully open model here. Data and code are public. Nothing up its sleeve.
OLMo 2 32B
Earlier fully open model. Brilliant and honest, but forgets what you said a few pages back.
Granite 4.0 H Small
Bigger IBM model for long documents. Wants about 26 GB.
Qwen3 30B-A3B 2507
Big on paper, quick in practice. Only part of it works on each word.
Qwen3 Coder 30B
Writes code so you do not have to. You still have to read it.
64 GB of memory
41 models fit. Here are the biggest ones that fit, since bigger usually means better answers.
Llama 3.3 70B
Smart, but wants about 53 GB of memory. Not a laptop guest.
Qwen3 32B
A careful thinker. Wants about 26 GB and a little patience.
Olmo 3.1 32B
Largest fully open model here. Data and code are public. Nothing up its sleeve.
OLMo 2 32B
Earlier fully open model. Brilliant and honest, but forgets what you said a few pages back.
Granite 4.0 H Small
Bigger IBM model for long documents. Wants about 26 GB.
Qwen3 30B-A3B 2507
Big on paper, quick in practice. Only part of it works on each word.
128 GB of memory
44 models fit. Here are the biggest ones that fit, since bigger usually means better answers.
GPT-OSS 120B
Hard reasoning and agent tasks on a high-end workstation. Plan on a 128 GB machine. Your laptop just flinched.
GLM-4.5 Air
Strong agent and coding work. It is called Air and it wants a 128 GB machine. Someone in naming has jokes.
Llama 4 Scout
Reads images too. Wants about 81 GB, so bring a big machine and snacks.
Llama 3.3 70B
Smart, but wants about 53 GB of memory. Not a laptop guest.
Qwen3 32B
A careful thinker. Wants about 26 GB and a little patience.
Olmo 3.1 32B
Largest fully open model here. Data and code are public. Nothing up its sleeve.
Buying guide
Want a computer for local AI? Here are four rough price tiers. Prices are the maker's own starting price, about, as of Oct 2026. Upgrades cost extra, so check the final price before you buy. The model picks use our RAM estimates.
About $600 to $1,000
The starter
A base Mac mini with 16 GB, or an older PC with a new 16 GB graphics card added.
- Mac mini (M5): 16 GB unified memory to start, up to 64 GB. From about $899, as of Oct 2026. Maker page, checked Oct 14, 2026.
- GeForce RTX 5060 Ti: 16 GB or 8 GB of VRAM. From about $379 for the card, as of Oct 2026. Maker page, checked Oct 14, 2026.
- GeForce RTX 5070: 12 GB of VRAM. From about $549 for the card, as of Oct 2026. Maker page, checked Oct 14, 2026.
Runs well: Mistral Nemo 12B, Gemma 3 12B, Gemma 4 12B, Qwen3.5 9B, Nemotron Nano 9B v2.
Runs the small and mid models well. Your first nerd party, but you get to leave early.
About $1,000 to $2,000
The laptop or gaming PC
A MacBook Air with 24 or 32 GB, or a gaming PC with a 16 GB NVIDIA card.
- MacBook Air (M5): 16 GB unified memory to start, up to 32 GB. From about $1,299, as of Oct 2026. Maker page, checked Oct 14, 2026.
- GeForce RTX 5070 Ti: 16 GB of VRAM. From about $749 for the card, as of Oct 2026. Maker page, checked Oct 14, 2026.
- GeForce RTX 5080: 16 GB of VRAM. From about $999 for the card, as of Oct 2026. Maker page, checked Oct 14, 2026.
Runs well: Qwen3 14B, DeepSeek R1 Distill 14B, Phi-4, Mistral Nemo 12B, Gemma 3 12B.
The gaming PC was for homework. Sure it was.
About $2,000 to $3,000
The serious desk
A MacBook Pro or Mac Studio with 32 GB or more, or a PC with a 32 GB RTX 5090.
- MacBook Pro (M5, M5 Pro, M5 Max): Up to 128 GB unified memory with M5 Max. From about $1,999, as of Oct 2026. Maker page, checked Oct 14, 2026.
- Mac Studio (M5 Max, M5 Ultra): 36 GB unified memory to start, up to 512 GB with M5 Ultra. From about $2,499, as of Oct 2026. Maker page, checked Oct 14, 2026.
- GeForce RTX 5090: 32 GB of VRAM. From about $1,999 for the card, as of Oct 2026. Maker page, checked Oct 14, 2026.
Runs well: Qwen3 32B, Olmo 3.1 32B, OLMo 2 32B, Granite 4.0 H Small, Qwen3 30B-A3B 2507.
This is where the 27B to 32B models stop being slow and start being fun.
Over about $2,500 (exact prices not yet checked)
The big boxes
A mini PC with AMD Ryzen AI Max+ and 128 GB, NVIDIA DGX Spark with 128 GB, or a Mac Studio set up with 128 GB or more.
- Mini PCs with AMD Ryzen AI Max+ 395: Up to 128 GB of memory, shared with the Radeon 8060S graphics. Price not yet checked. Maker page, checked Oct 14, 2026.
- NVIDIA DGX Spark: 128 GB of unified memory (a 64 GB option is also listed). Price not yet checked. Maker page, checked Oct 14, 2026.
- Mac Studio (M5 Max, M5 Ultra): 36 GB unified memory to start, up to 512 GB with M5 Ultra. From about $2,499, as of Oct 2026. Maker page, checked Oct 14, 2026.
Runs well: GPT-OSS 120B, GLM-4.5 Air, Llama 4 Scout, Llama 3.3 70B, Qwen3 32B.
Big enough for the 120B models. The nerd club will want to borrow it.
Graphics card prices are for the card only, not a whole PC. On a PC, the model needs to fit in the card's VRAM to run fast. On a Mac or a Ryzen AI Max mini PC, the memory is shared, so the full memory number counts.
Can my kid's gaming PC do this?
Maybe! Here is a two-minute check. No nerd club card needed.
- On Windows, right-click the taskbar and open Task Manager.
- Click Performance, then GPU. The name at the top is the graphics card.
- Look for Dedicated GPU memory. That is the VRAM.
- 8 GB of VRAM runs small models. 12 to 16 GB runs the good everyday ones. 24 GB or more runs the strong ones.
On a Mac, open the Apple menu and pick About This Mac. The Memory line is your number.
Common questions
How much RAM do I need to run local AI?
About 8 GB runs small models. 16 GB runs good everyday models. 32 GB or more runs strong ones.
Is a Mac good for local AI?
Yes. Macs with Apple chips share memory with the graphics, so a 32 GB Mac can run models a 32 GB PC without a big graphics card would find slow.