Running an AI model on your own computer keeps your prompts, documents and code on the machine, and you stop paying by the month or by the token. Free apps like LM Studio and Ollama download an open model and run it in a chat window.
Your machine's memory limits what you can run. A model has to fit in memory before it starts, and the popular 4-bit files on Hugging Face show the scale: an 8-billion-parameter model is 4.9 GB, a 32B model is 19.8 GB and a 70B model is 42.5 GB. Memory bandwidth comes second, because it sets how fast the answer appears.
Apple refreshed the Mac mini and Mac Studio on August 25, with orders arriving September 22, and the 2026 memory shortage has raised the price of every machine with a lot of RAM.
Apple Mac mini (2026, M6, 24 GB)
M6 · 24 GB unified memory · 170 GB/s · 512 GB SSD · ships September 22
The Mac mini is a small, quiet desktop that holds any 8B or 14B model with room left for your apps. Those sizes handle everyday chat, summaries and writing help.
- Holds 8B and 14B models with room to spare
- Small and silent, and runs LM Studio and Ollama natively
- Amazon sells it directly
- A 32B model's 19.8 GB file leaves too little of 24 GB for macOS
- 170 GB/s is the slowest memory here
- New model, so no reviews yet
Apple's M6 Mac mini starts at $899 with 16 GB, and this 24 GB, 512 GB version is $1,299 on Amazon. The extra 8 GB lets a 14B model, a 9.0 GB file at 4-bit, run while your browser and editor stay open. Apple quotes 170 GB/s of memory bandwidth for M6, the lowest here, which suits small models and drags on anything near the memory limit. Amazon lists it for September 22 delivery.
Apple Mac Studio (2026, M5 Max, 36 GB)
M5 Max · 36 GB unified memory · 614 GB/s · ships September 22
The M5 Max Mac Studio is the fastest machine here for models up to 32B. A 32B model's 19.8 GB file fits in 36 GB, and the M5 Max has the fastest memory of any pick.
- Fastest memory here at 614 GB/s
- Room for 32B models with memory left over
- Configurable to 128 GB through Apple
- Amazon lists only the 36 GB version
- 70B models don't fit in 36 GB
- Costs about twice the Mac mini
The M5 Max Mac Studio starts at $2,499 at Apple and $2,449.99 on Amazon, where only the 36 GB base model is listed. Apple rates M5 Max at up to 614 GB/s of memory bandwidth, more than twice what AMD's 128 GB mini PCs have, and bandwidth sets how many words per second you get once a model is loaded. If you already know you want 70B models, configure it with 128 GB on Apple's site, since Amazon doesn't carry that version.
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB / 2 TB)
Ryzen AI Max+ 395 · 128 GB LPDDR5X-8000 · up to 96 GB as graphics memory · 2 TB SSD
The EVO-X2 is a small desktop with 128 GB of memory for 70B models. AMD's Ryzen AI Max+ 395 shares that memory between processor and graphics, and AMD lets up to 96 GB of it work as graphics memory, enough for a 70B model's 42.5 GB file with room for long conversations.
- 128 GB fits 70B models with room to spare
- Runs LM Studio, which GMKtec's listing calls out
- Also a full desktop PC for everything else
- 256 GB/s of bandwidth, so 70B models answer slowly
- Sold by GMKtec, not Amazon
- 4.2 stars across 111 ratings
AMD's spec sheet lists a 256-bit LPDDR5x-8000 memory bus for this chip, which works out to 256 GB/s. That is less than half the M5 Max's bandwidth, so the EVO-X2 holds models the 36 GB Mac can't, then runs them more slowly. The Beelink GTR9 Pro and Minisforum MS-S1 Max use the same chip and cost $3,799 to $4,349 on Amazon, which makes the EVO-X2 the value version.
ASUS ROG Flow Z13 (Ryzen AI Max+ 395, 128 GB)
Ryzen AI Max+ 395 · 128 GB LPDDR5X-8000 · 13.4" 2.5K 180 Hz touchscreen · detachable keyboard · Windows 11 Pro
The Flow Z13 puts the same chip and 128 GB as the EVO-X2 in a 13-inch tablet with a detachable keyboard. It costs $350 less than the EVO-X2 on Amazon, and Amazon sells it directly.
- 70B models on a machine you can carry
- Sold and shipped by Amazon
- Kickstand and touchscreen for tablet use
- Same 256 GB/s bandwidth, so 70B answers come slowly
- 1 TB fills fast when one 70B model is 42.5 GB
- A 13-inch screen is small for long work sessions
ASUS puts AMD's Ryzen AI Max+ 395 and 128 GB of LPDDR5X-8000 in a 13.4-inch tablet with a detachable keyboard and a 170-degree kickstand. It runs the same models as the EVO-X2 at the same speed, and at $3,299.99 it undercuts every 128 GB mini PC on Amazon. If you plan to keep several large models, budget for an external SSD.
ASUS Prime GeForce RTX 5060 Ti 16GB
16 GB GDDR7 · 448 GB/s · 128-bit bus · PCIe 5.0 · 180 W
The RTX 5060 Ti goes in the desktop you already own. Most image and video generation tools, ComfyUI included, support NVIDIA cards without extra setup, and this card brings 16 GB of fast graphics memory for about $790.
- 16 GB of GDDR7 at 448 GB/s
- CUDA support for image, video and language-model tools
- Sold by Amazon, 4.7 stars across 167 ratings
- About $790, far above its $429 launch price
- Needs a desktop with a free slot and power for a 180 W card
- Language models stop at about 14B in 16 GB
NVIDIA launched the 16 GB RTX 5060 Ti at $429 in April 2025, and it now sells for about $790, part of a memory shortage that hit GDDR7 cards hardest. It is still the cheapest current NVIDIA card with 16 GB. For language models it runs 8B and 14B fast, since 448 GB/s beats every machine here except the Mac Studio, but anything past 14B spills out of its memory.
The numbers.
| Mac mini M6 | Mac Studio M5 Max | GMKtec EVO-X2 | ROG Flow Z13 | RTX 5060 Ti | |
|---|---|---|---|---|---|
| Best for | Small models | Mid-size models | 70B at a desk | 70B on the go | Image generation |
| Memory | 24 GB | 36 GB | 128 GB | 128 GB | 16 GB on the card |
| Bandwidth | 170 GB/s | 614 GB/s | 256 GB/s | 256 GB/s | 448 GB/s |
| Fits up to | 14B | 32B | 70B | 70B | 14B |
| Form | Mini desktop | Desktop | Mini desktop | 13" tablet | PCIe card |
| Price | ~$1,299 | ~$2,450 | ~$3,650 | ~$3,300 | ~$790 |
Bigger, newer, or skip.
Apple Mac Studio (M5 Ultra)
The M5 Ultra Mac Studio is the Mac for models bigger than 128 GB can hold. Apple's M5 Ultra starts at $5,499 and configures up to 512 GB of unified memory with 1.2 TB/s of bandwidth, per Apple. Orders arrive from September 22, and the 512 GB version follows in late October. Amazon doesn't list the M5 Ultra, so order it from Apple.
Visit apple.com →NVIDIA DGX Spark
NVIDIA builds this 128 GB desktop on its own GB10 chip, with 273 GB/s of memory, for people who want NVIDIA's AI software on a desk. NVIDIA raised the price from $3,999 to $4,699 in February 2026, citing memory supply. On Amazon, ASUS's version sells through third-party resellers for about $6,000, so buy it from NVIDIA.
Visit marketplace.nvidia.com →NVIDIA RTX Spark PCs (October 2026)
NVIDIA says the first Windows laptops and compact desktops with its RTX Spark chip arrive in October from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, and hasn't announced prices. If you want a Windows laptop for 70B models and can wait a few weeks, see what they cost before you buy the ROG Flow Z13.
GeForce RTX 5090 (skip for now)
The RTX 5090 is NVIDIA's fastest consumer card, with 32 GB of GDDR7. It lists for about $6,000 on Amazon, and even 32 GB can't hold a 70B model's 42.5 GB file. A 128 GB mini PC runs bigger language models for a little over half the money.
The buying guide.
How much memory each model needs
A model has to fit in memory before it runs at all. The common 4-bit files on Hugging Face give the sizes: Llama 3.1 8B is 4.9 GB, Qwen3 14B is 9.0 GB, Qwen3 32B is 19.8 GB and Llama 3.3 70B is 42.5 GB. Add a few gigabytes for the conversation and for your operating system. So 24 GB handles 14B, 36 GB handles 32B, and 70B needs 64 GB at minimum and 128 GB to be comfortable.
Bandwidth sets the speed
For every token it writes, the computer reads the whole model out of memory, so bandwidth divided by model size sets a ceiling on speed. The Mac Studio's 614 GB/s over a 19.8 GB model works out to about 31 tokens a second at most. The EVO-X2's 256 GB/s over a 42.5 GB model tops out near 6. Real speeds land below those ceilings, but they rank the machines in the same order.
Mac or Windows
Both run LM Studio and Ollama, the two free apps most people start with. A Mac uses one pool of memory for everything, so a model can use most of it. On Windows, a regular graphics card only uses its own memory, which is why a 16 GB card stops at 14B models while AMD's Ryzen AI Max+ chips can hand up to 96 GB to the graphics. For image and video generation, an NVIDIA card is the path with the fewest setup steps.
Why these machines cost more in 2026
Every machine here sells on its memory, and memory got scarce this year. NVIDIA raised the DGX Spark from $3,999 to $4,699 in February, citing memory supply, and the RTX 5060 Ti sells for nearly double its launch price. Our RAM price explainer covers why, and our earlier RTX Spark piece covers the machines NVIDIA announced in June.
FAQ.
One with enough memory for the model you want. A 16 GB or 24 GB machine runs 8B and 14B models, 36 GB runs 32B models, and 128 GB runs 70B models. Memory bandwidth then sets how fast the answers come.
A 70B model at 4-bit is about 42.5 GB, so plan on 64 GB at minimum and 128 GB for long conversations. On a Windows PC with a regular graphics card, the model has to fit in the card's own memory, which rules out 70B on any single consumer card.
For language models, a Mac shares one fast pool of memory, and the Mac Studio has the fastest memory of any machine here. For image and video generation, a PC with an NVIDIA card is the easier setup. AMD's 128 GB mini PCs sit in between: lots of memory, slower bandwidth.
Yes, if it has an NVIDIA card with 12 GB or more of memory. That runs 8B and 14B models and image generation. If your card has less, an RTX 5060 Ti 16GB is the cheapest current NVIDIA upgrade to 16 GB.
If you want a Windows laptop for large models and can wait until October, wait for the prices, since NVIDIA hasn't announced any. For a desktop, today's 128 GB machines already run the same size of models.
Buy the memory
for the model you want.
For chat, writing help and summaries, the 24 GB Mac mini M6 is the easy start at $1,299. For 32B models at speed, the Mac Studio M5 Max. For 70B, the GMKtec EVO-X2 on a desk or the ROG Flow Z13 in a bag, both with 128 GB. If images are the goal, put an RTX 5060 Ti in the PC you already own.