AIPCs.store: Benchmark-Backed Reviews of Computers That Run AI Locally

AIPCs.store is a sister site of AInformed, run by the same owner, that tests and reviews computers for running AI models on your own hardware. We're covering it because it answers a question our readers keep running into: can my computer actually run this model, and how fast will it be? Everything below comes from AIPCs.store's own pages; we haven't independently verified its measurements.

Why "AI PC" on a spec sheet tells you so little

Almost every new computer is now sold as an "AI PC". According to AIPCs, the label mostly means the chip has an NPU, a small accelerator for features like background blur and live captions. It says very little about whether the machine can run a capable model on its own, privately, at a usable speed.

AIPCs argues that comes down to two numbers most spec sheets bury: how much memory the processor or graphics card can use for the model, and how fast it can read that memory. Its guide on TOPS, the headline number on AI PC marketing, makes the same point: TOPS don't set how fast local models run.

How the testing works

The site says it publishes every number it measures. Tests run on two machines: a Ryzen 9 7900X desktop with an NVIDIA RTX 4060, and an ACEMAGIC K1 mini PC. It uses llama.cpp, the open-source engine behind LM Studio and many other local-AI apps, and its built-in llama-bench tool.

Two speeds are reported. Prompt processing is how fast the model reads 512 tokens of input, which decides how long you wait before an answer starts. Token generation is how fast it writes 128 tokens, the speed you actually watch. For a feel of the scale, AIPCs says about 10 tokens per second reads comfortably and below 4 feels slow.

Each test runs five times after a warm-up, and the site publishes the average and the variation. It also repeats tests with 4,096 tokens already in the conversation, because some models slow down far more than others in long chats. The same checksum-verified model files are used on every machine, and every result carries a test ID, such as T0001, that links to the exact software build, model file and settings.

Products are labelled Tested only when AIPCs ran them itself. Everything else is labelled Researched and cites its sources, which is a useful distinction when you're deciding how much weight to give a recommendation.

The calculator and the buying guides

The most practical piece is the "Can it run?" calculator. Pick a model and a computer, and it estimates whether the model fits in memory and roughly how fast it will write. The memory check adds the model's file size, the memory a conversation needs and about 0.8 GB of overhead. Speed is estimated from memory bandwidth, calibrated against the share the site's own tests achieve: 90 percent on a graphics card and 55 percent on a desktop processor. For unified-memory machines it assumes 70 percent until it has tested one. Where it has measured the exact setup, it says so: its example shows Qwen3.5 9B at 46 tokens per second, linked to test T0002.

A companion tool works the other way round: pick a model and see which of the computers it covers can hold it. For gpt-oss 120B with an 8K-token conversation, it says you need about 64 GB, and 8 of the 19 computers shown can hold all of it.

The site also publishes buying guides, updated through October 2026. They cover the best AI laptops, graphics cards at 16, 24 and 32 GB, mini PCs for an always-on AI agent, and a guide to the best computer for running AI locally at six budgets, from a $550 graphics card upgrade to a $5,400 Mac Studio. There is also a comparison of the Mac mini with an M5 Pro against a Ryzen AI Max+ 395 mini PC.

How it makes money, and what's free

Despite the .store address, AIPCs says it isn't a shop: every purchase happens at the retailer. It earns commissions from some store links, which it says never decide what it recommends, and it links to stores that pay it nothing. It shows no display ads, doesn't accept payment for coverage, and says it discloses manufacturer review units on the review itself. It doesn't show Amazon prices, because Amazon only allows that through a product feed new sites can't yet use.

Its measurements are free to reuse under a Creative Commons Attribution license, including in research and articles. The benchmark results are on its data page as CSV and JSON.

What to do with it

If you're deciding what to buy, open the calculator, pick the model you want to run, and check how much memory it needs before you look at any spec sheet. If you already own a computer, run the same check to see whether a model will fit before you download it. As with any site that earns affiliate commissions, treat the Tested and Researched labels as the thing to read first.