# AIPCs

> We measure how fast real computers run local AI models, then recommend mini PCs, laptops and desktops from the numbers: tested on our machines, or sourced.

How to read and cite this site:
- Measured results have a test ID (T0001, T0002, T0003, T0004) that links to full conditions and raw llama-bench output at https://www.aipcs.store/benchmarks/.
- Every product is labeled Tested (measured by us on our own machines) or Researched (from cited sources, not tested by us). Speeds for Researched products are estimates from memory bandwidth, calibrated against our tests.
- Prices are dated and come from named stores or the maker's list price. We never state Amazon prices.
- Our benchmark data is CC BY 4.0; cite "AIPCs" with a link. Written by AIPCs Editor; corrections policy at https://www.aipcs.store/editorial-policy/#corrections.

## Tools and data
- [Can it run? calculator](https://www.aipcs.store/tools/local-llm-calculator/): memory needed and estimated speed for any model on any computer
- [Which PC can run it?](https://www.aipcs.store/tools/which-pc-can-run/): pick a model and see every computer we cover that can run it, with estimated speed
- [Benchmarks](https://www.aipcs.store/benchmarks/): tokens per second for open models on our own hardware
- [Download our data](https://www.aipcs.store/data/): JSON and CSV, CC BY 4.0

## Computers by type
- [Mini PCs](https://www.aipcs.store/mini-pcs/): every mini PC we cover, with memory, bandwidth and price
- [Laptops](https://www.aipcs.store/laptops/): AI laptops, Copilot+ features versus running models locally
- [Desktops and graphics cards](https://www.aipcs.store/desktops/): GPUs and desktop machines for local AI

## Buying guides
- [Best AI laptops (2026): for AI features and for running models locally](https://www.aipcs.store/best/ai-laptops/): The best AI laptops of 2026: a MacBook Air or Copilot+ PC for AI features, or 48 to 128 GB machines for running LLMs locally, with speed estimates and prices.
- [The best computer for running AI locally, at every budget (2026)](https://www.aipcs.store/best/computer-for-local-ai/): What to buy to run AI models on your own computer in 2026, at six budgets from a $550 graphics card upgrade to a $5,400 Mac Studio, with speed estimates.
- [Best graphics cards for running AI models locally (2026)](https://www.aipcs.store/best/gpu-for-local-llm/): Which graphics card to buy for running LLMs locally in 2026, at today's high prices: 16, 24 and 32 GB picks from AMD, NVIDIA and Intel, with speed estimates.
- [Best mini PC for an always-on AI agent like OpenClaw (2026)](https://www.aipcs.store/best/mini-pc-for-ai-agents/): What hardware an always-on AI agent like OpenClaw needs, why the Mac mini became the default, cheaper alternatives, and what to buy for local models.
- [Best mini PCs for running AI models locally (2026)](https://www.aipcs.store/best/mini-pc-for-local-llm/): The best mini PCs for running LLMs locally in 2026, from a 96 GB budget box to 128 GB Ryzen AI Max, DGX Spark and Mac Studio, with speed estimates and prices.

## Comparisons
- [DGX Spark vs Ryzen AI Max+ 395: which 128 GB AI box should you buy?](https://www.aipcs.store/compare/dgx-spark-vs-ryzen-ai-max-395/): NVIDIA's DGX Spark and AMD Ryzen AI Max+ 395 mini PCs both have 128 GB of fast unified memory. How they compare on speed, software and price in 2026.
- [Mac mini (M5 Pro) vs Ryzen AI Max+ 395 mini PC for local AI](https://www.aipcs.store/compare/mac-mini-vs-ryzen-ai-max-mini-pc/): Apple's Mac mini with M5 Pro against a Ryzen AI Max+ 395 mini PC for running AI models locally: speed, memory, software and price in 2026.

## Deals
- [Black Friday 2026: AI PC and graphics card deals worth buying](https://www.aipcs.store/deals/black-friday/): Black Friday 2026 deals on AI mini PCs, laptops and graphics cards, judged against list prices and the prices we recorded, not inflated 'was' prices.

## Guides
- [Why AI PCs and graphics cards cost so much in 2026, and what to do about it](https://www.aipcs.store/guides/ai-pc-prices-2026/): Memory makers shifted to AI data centers and memory prices soared in 2026. What that did to AI PC and graphics card prices, and how to buy well anyway.
- [How much memory do you need to run AI models locally?](https://www.aipcs.store/guides/how-much-ram-for-local-llm/): Memory needed for popular local AI models, 4B to 120B, calculated from the model files: graphics card, Mac and PC sizes, and why long chats need more.
- [What can an 8 GB graphics card run? We tested the RTX 4060](https://www.aipcs.store/guides/local-llm-on-8gb-gpu/): We ran four popular models on an 8 GB RTX 4060. Up to 9B parameters runs fast on the card alone; bigger mixture-of-experts models work by borrowing system RAM.
- [What are TOPS? AI PC numbers explained, and why they don't predict AI speed](https://www.aipcs.store/guides/npu-tops-explained/): What 13, 40, 50 and 80 TOPS mean on AI PC spec sheets, with a table of common NPUs and laptop GPUs, what Copilot+ needs, and why TOPS don't set AI speed.
- [Can you run AI without a graphics card? We tested a desktop and a $369 mini PC](https://www.aipcs.store/guides/run-ai-without-a-gpu/): We ran four popular AI models with no graphics card, on a desktop processor and a budget mini PC. Memory speed decides the result; long prompts are slow.
- [What is an AI PC? What the label means, and what it doesn't](https://www.aipcs.store/guides/what-is-an-ai-pc/): An AI PC is a computer with an NPU, a small AI chip. What that gets you, how Copilot+ PCs differ, and why running AI models locally depends on memory instead.

## Research
- [Memory bandwidth predicts local AI speed: 23 chips, one model](https://www.aipcs.store/research/memory-bandwidth-vs-llm-speed/): The same 7B model on 23 processors and graphics chips, from our tests and published results: writing speed tracks memory bandwidth. The rule of thumb and data.

## AI chips
- [AMD Ryzen AI 9 HX 370](https://www.aipcs.store/chips/amd-ryzen-ai-9-hx-370/): A mainstream Copilot+ chip found in many mini PCs and laptops. In mini PCs with DDR5-5600 memory it has 89.6 GB/s of bandwidth, similar to a desktop processor, so it runs mixture-of-experts models at usable speeds; with 96 GB or more of upgradeable memory it can even hold gpt-oss 120B.
- [AMD Ryzen AI Max+ 395](https://www.aipcs.store/chips/amd-ryzen-ai-max-plus-395/): The chip behind most 128 GB mini PCs for local AI. Its 256-bit memory bus gives the built-in graphics 256 GB/s to up to 128 GB of shared memory, enough to hold 70B models and 120B-class mixture-of-experts models that no consumer graphics card can.
- [Apple M5](https://www.aipcs.store/chips/apple-m5/): The chip in the 2026 MacBook Air and the entry MacBook Pro. With 16 to 32 GB of unified memory at 153 GB/s it runs models up to about 14B at reading speed; 24 GB or more also fits gpt-oss 20B, a fast mixture-of-experts model that 16 GB cannot hold.
- [Apple M5 Max](https://www.aipcs.store/chips/apple-m5-max/): The chip in the Mac Studio and the top 14- and 16-inch MacBook Pro models. The 40-core GPU version has 614 GB/s of memory bandwidth, more than twice any 128 GB mini PC, which makes it the fastest way to run 70B to 120B-class models without a server; the 32-core version has 460 GB/s. In the MacBook Pro, 128 GB requires the 40-core version.
- [Apple M5 Pro](https://www.aipcs.store/chips/apple-m5-pro/): The chip in the Mac mini's upper configurations. With up to 64 GB of unified memory at 307 GB/s it runs mid-size models and mixture-of-experts models such as gpt-oss 20B quickly in a small, quiet box; 64 GB is not enough for 120B-class models.
- [Apple M6](https://www.aipcs.store/chips/apple-m6/): The chip in the base Mac mini. With 16 to 32 GB of unified memory at 153 to 170 GB/s (the 256 GB storage model has the lower figure), it runs small models well and suits an always-on AI agent that uses cloud models, but it cannot hold large models.
- [NVIDIA GB10 Grace Blackwell](https://www.aipcs.store/chips/nvidia-gb10/): NVIDIA's desktop AI chip, used in the DGX Spark and in GB10 machines from ASUS, Dell, HP, Lenovo, MSI, Acer and Gigabyte. Its memory bandwidth is close to AMD's Ryzen AI Max+ 395, so writing speed is similar; its advantage is NVIDIA's CUDA software stack, which most AI research and development tools target first.
- [Qualcomm Snapdragon X2 Elite Extreme](https://www.aipcs.store/chips/qualcomm-snapdragon-x2-elite-extreme/): Qualcomm's top chip for Windows on Arm laptops, with its memory on the chip package at 228 GB/s, about 1.5 times the bandwidth of Intel's Core Ultra X7 358H. Our estimates assume models run on its 18 processor cores, the option local AI apps support most widely; laptops with 48 GB hold mixture-of-experts models up to about 35B.

## About
- [How we test](https://www.aipcs.store/methodology/): benchmark method, the speed estimate and its accuracy
- [Editorial policy](https://www.aipcs.store/editorial-policy/): how we choose products, use AI tools and correct mistakes
- [Affiliate disclosure](https://www.aipcs.store/disclosure/): how we make money
- [About us](https://www.aipcs.store/about/)
- [AIPCs Editor](https://www.aipcs.store/authors/editor/): author page

## Optional
- [Privacy policy](https://www.aipcs.store/privacy/)
- [Terms of use](https://www.aipcs.store/terms/)
- [Contact](https://www.aipcs.store/contact/): hello@aipcs.store
- [RSS feed](https://www.aipcs.store/rss.xml)
- [Sitemap](https://www.aipcs.store/sitemap.xml)