Local AI hardware, measured in models
Find the hardware that runs the model you want
Pick a model and we show every GPU, Mac, mini PC and edge board that can hold it, how fast it can go and what it costs. No hype, just memory and math.
What do you want to run?
33 platforms · 21 models · prices checked Oct 12, 2026 · Read the methodology
Seven ways to run an LLM at home
View all →Speed depends on memory bandwidth, model size on memory capacity. Each family trades one for the other differently.
- 12 platforms · 12 GB–96 GBNVIDIA graphics cardsThe fastest option and CUDA, which every AI tool supports first.Memory per card is small (12–32 GB) and 2026 prices are inflated.
- 4 platforms · 16 GB–32 GBAMD Radeon graphics cardsMore memory per euro, including a 32 GB card, with ROCm and Vulkan.Software is good with llama.cpp and Ollama, thinner elsewhere.
- 4 platforms · 12 GB–32 GBIntel Arc graphics cardsThe cheapest new 24–32 GB cards, plus a 70 W card for tiny servers.The software path (SYCL, Vulkan) needs more tinkering.
- 4 platforms · 32 GB–512 GBApple silicon MacsUp to 512 GB of fast unified memory in a quiet box. MLX and llama.cpp run great.Memory upgrades are expensive and cannot be added later.
- 3 platforms · 96 GB–192 GBAMD Ryzen AI mini PCs and NASRyzen AI Max: 128–192 GB of unified memory in an x86 mini PC or NAS. Ryzen AI 9: cheap boxes up to 96 GB.Max runs at about 256–273 GB/s and Ryzen AI 9 at 90–120 GB/s: dense models over 30B (or 14B) feel slow.
- 1 platforms · 128 GB–128 GBNVIDIA DGX Spark and GB10 systems128 GB with the full CUDA stack in a tiny box.Same bandwidth class as Strix Halo for a higher price; Arm CPU.
- 5 platforms · 8 GB–128 GBEdge AI boardsJetson, Raspberry Pi AI HAT+ 2 and RK3588: small models at a few watts.Small models only; some run vendor-compiled models only.
Memory vs speed: the whole market in one chart
Further right holds bigger models; higher generates text faster. Graphics cards are fast but small; unified-memory machines are big but slower.
- Graphics card
- Unified-memory computer
- Edge board / accelerator
Show the chart as a table
| Platform | Type | Memory | Bandwidth |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell | Graphics card | 96 GB | 1.79 TB/s |
| NVIDIA GeForce RTX 5090 | Graphics card | 32 GB | 1.79 TB/s |
| NVIDIA RTX PRO 5000 Blackwell | Graphics card | 48 GB | 1.34 TB/s |
| Apple Mac Studio (M5 Ultra) | Unified-memory computer | 512 GB | 1.2 TB/s |
| NVIDIA GeForce RTX 4090 | Graphics card | 24 GB | 1.01 TB/s |
| AMD Radeon RX 7900 XTX | Graphics card | 24 GB | 960 GB/s |
| NVIDIA GeForce RTX 5080 | Graphics card | 16 GB | 960 GB/s |
| NVIDIA GeForce RTX 3090 | Graphics card | 24 GB | 936 GB/s |
| NVIDIA RTX PRO 4500 Blackwell | Graphics card | 32 GB | 896 GB/s |
| NVIDIA GeForce RTX 5070 Ti | Graphics card | 16 GB | 896 GB/s |
| NVIDIA RTX PRO 4000 Blackwell | Graphics card | 24 GB | 672 GB/s |
| NVIDIA GeForce RTX 5070 | Graphics card | 12 GB | 672 GB/s |
| AMD Radeon AI PRO R9700 | Graphics card | 32 GB | 645 GB/s |
| AMD Radeon RX 9070 XT | Graphics card | 16 GB | 645 GB/s |
| Apple Mac Studio (M5 Max) | Unified-memory computer | 128 GB | 614 GB/s |
| Intel Arc Pro B70 | Graphics card | 32 GB | 608 GB/s |
| Intel Arc Pro B60 | Graphics card | 24 GB | 456 GB/s |
| Intel Arc B580 | Graphics card | 12 GB | 456 GB/s |
| NVIDIA GeForce RTX 5060 Ti 16 GB | Graphics card | 16 GB | 448 GB/s |
| NVIDIA GeForce RTX 3060 12 GB | Graphics card | 12 GB | 360 GB/s |
| AMD Radeon RX 9060 XT 16 GB | Graphics card | 16 GB | 320 GB/s |
| Apple Mac mini (M5 Pro) | Unified-memory computer | 64 GB | 307 GB/s |
| AMD Ryzen AI Max+ PRO 495 (Gorgon Halo) | Unified-memory computer | 192 GB | 273 GB/s |
| NVIDIA Jetson AGX Thor Developer Kit | Edge board / accelerator | 128 GB | 273 GB/s |
| NVIDIA GB10 Grace Blackwell (DGX Spark and partners) | Unified-memory computer | 128 GB | 273 GB/s |
| AMD Ryzen AI Max+ 395 (Strix Halo) | Unified-memory computer | 128 GB | 256 GB/s |
| Intel Arc Pro B50 | Graphics card | 16 GB | 224 GB/s |
| NVIDIA Jetson AGX Orin 64GB Developer Kit | Edge board / accelerator | 64 GB | 205 GB/s |
| Apple Mac mini (M6) | Unified-memory computer | 32 GB | 170 GB/s |
| AMD Ryzen AI 9 HX 370 (Strix Point) | Unified-memory computer | 96 GB | 120 GB/s |
| NVIDIA Jetson Orin Nano Super Developer Kit | Edge board / accelerator | 8 GB | 102 GB/s |
Best price per GB of memory
Typical street price of the cheapest configuration divided by its memory. Lower is better.
Our picks by budget and use
View all →- Best graphics cards for local LLMsRTX 3090 · Radeon AI PRO R9700 · RTX 5090
- Best hardware to run gpt-oss-120b locallyRyzen AI Max+ 395 · Mac Studio M5 Max · DGX Spark / GB10
- Best budget hardware for local LLMsRTX 5060 Ti 16 GB · RX 9060 XT 16 GB · RTX 3060 12 GB
- Best mini PCs and small machines for local LLMsRyzen AI Max+ 395 · Mac Studio M5 Max · Mac mini M5 Pro
Popular models and what they need
View all →Latest guides
View all →- Comparisons· Oct 12, 2026Can a NAS run LLMs? The Ryzen AI NAS boxes explained
Two Minisforum NAS models carry Ryzen AI chips. One has the same chip as a 128 GB Strix Halo mini PC; the other is a mid-range box. What each can actually run.
- Comparisons· Oct 11, 2026Unified memory or a graphics card? Mac Studio, Strix Halo and DGX Spark vs GPUs
Why a 128 GB mini PC runs models a 32 GB GPU cannot, why the GPU is still faster, and how to choose for your homelab.
- Edge AI· Oct 11, 2026Can the Raspberry Pi AI HAT+ 2 run LLMs?
What the Hailo-10H accelerator can and cannot run, and when a Jetson or a small GPU is the better choice.
- Basics· Oct 11, 2026How much memory do you need to run an LLM locally?
Weights, context cache and overhead: how to work out whether a model fits your GPU or Mac, with real numbers for popular models.
Popular comparisons
View all →How we work
Memory needs come from real model files; speeds are a theoretical ceiling unless marked as measured, with the source. Prices are dated references.
Read the methodology →