Ollama makes running a local LLM a one-line job, but the box underneath decides what you can run. Two numbers matter more than anything else on the spec sheet: how much memory the GPU can use, which decides which models fit, and how fast that memory is, which decides how many tokens per second you get.
Our verdict: a Ryzen AI Max+ 395 "Strix Halo" mini PC with 128GB, led by the Framework Desktop at $3,449, is the best mini PC for Ollama in 2026. It gives you 128GB of fast unified memory for the money, enough for 100B-class models a Mac mini can't load. The Mac mini with M5 Pro and 64GB ($3,199) is the easiest pick and the fastest on models that fit in 64GB.
The Short Version
Best overall: Framework Desktop with the Ryzen AI Max+ 395 and 128GB, $3,449 for the DIY Edition. If it's out of stock, the GMKtec EVO-X2 128GB is $3,499.99.
Best Mac and easiest setup: Mac mini with M5 Pro and 64GB, $3,199. The fastest memory here at 307GB/s, and Ollama just works.
Best for CUDA developers: Nvidia DGX Spark. Nvidia's full software stack in a mini box, but now the most expensive way to get 128GB.
Best budget: Mac mini M6 with 24GB, $1,299, for 8B to 14B models. Skip the 16GB model for local AI.
Prices at a glance
Prices checked October 15, 2026, on each company's own pricing page.
| Mini PC | Price | Max memory | Memory bandwidth | Ollama GPU support |
|---|---|---|---|---|
| Framework Desktop (Max+ 395) | $1,959 (64GB), $3,449 (128GB), before storage | 128GB | 256GB/s | ROCm (Linux) or Vulkan |
| GMKtec EVO-X2 (Max+ 395) | $2,199.99 (64GB), $3,499.99 (128GB, 1TB) | 128GB | 256GB/s | ROCm (Linux) or Vulkan |
| Beelink GTR9 Pro (Max+ 395) | $4,349 (128GB, 2TB) | 128GB | 256GB/s | ROCm (Linux) or Vulkan |
| Mac mini M5 Pro | $1,699 (24GB), $3,199 (64GB, 1TB) | 64GB | 307GB/s | Metal |
| Mac mini M6 | $899 (16GB), $1,299 (24GB), $1,799 (32GB, 1TB) | 32GB | 153 to 170GB/s | Metal |
| Nvidia DGX Spark | 64GB from $4,999 (partners, Oct. 23); 128GB price not listed | 128GB | 273GB/s | CUDA |
| Beelink SER9 Max (Ryzen 7 H 255) | $1,149 (64GB DDR5-5600) | 64GB | About 90GB/s | Vulkan or CPU |
Why memory decides everything
A model has to sit in GPU-accessible memory to run fast. A 4-bit 27B model like Qwen3.8 27B needs about 18GB, so 24GB is tight once the operating system takes its share and 32GB is comfortable. 100B-class models such as Qwen 3.5 122B-A10B need roughly 70GB or more, which means 96GB to 128GB. Our RAM guide for local LLMs goes model by model.
Speed comes from bandwidth. Each generated token reads the model's active weights from memory, so a rough ceiling is bandwidth divided by the gigabytes read per token. An 18GB dense model tops out near 17 tokens per second at 307GB/s, near 14 at 256GB/s, and about 5 at 90GB/s. Mixture-of-experts models read only their active experts, which is why 120B-class MoE models run fast on 128GB boxes.
Framework Desktop: the best mini PC for Ollama
AMD's Ryzen AI Max+ 395 pairs a 16-core CPU with a 40-core Radeon GPU on a 256-bit LPDDR5X-8000 bus, or 256GB/s, and lets the GPU use most of up to 128GB of shared memory: Mac-style unified memory at a lower price per gigabyte.
Framework's version is the one to buy. The DIY Edition is $1,959 with 64GB or $3,449 with 128GB, and you add your own SSD and operating system. Framework's own table shows the 128GB system running a 2-bit DeepSeek V4 Flash in 90.9GB. Ollama supports the chip through ROCm v7 on Linux and through its Vulkan backend on Windows and Linux.
The catch: every configuration was out of stock when we checked, and the memory is soldered, so choose 128GB up front.
GMKtec EVO-X2 and Beelink GTR9 Pro: same chip, different deal
The GMKtec EVO-X2 uses the same Ryzen AI Max+ 395 with LPDDR5X-8000. It's $2,199.99 with 64GB, $3,499.99 with 128GB and a 1TB SSD, or $3,649.99 with 2TB, and it arrives ready to use.
The Beelink GTR9 Pro comes only as 128GB with a 2TB SSD at $4,349, as a pre-order with about 35 days' wait. It takes up to 16TB of storage, but you're paying about $700 more than the 2TB EVO-X2 for the same memory and bandwidth.
The catch: both have soldered memory and the same 256GB/s ceiling as the Framework. GMKtec's page lists storage options that don't match its spec section, so confirm the SSD before ordering.
Mac mini: the easiest way to run Ollama
Ollama runs on Apple Silicon through Metal with no drivers to install. The M5 Pro model has 307GB/s of bandwidth, the fastest of any mini PC here. It's $1,699 with 24GB or $3,199 for the 18-core CPU, 20-core GPU version with 64GB and 1TB. The M6 model is $899 with 16GB, $1,299 with 24GB and $1,799 with 32GB and 1TB, at 153GB/s to 170GB/s.
On the llama.cpp project's long-running Apple Silicon benchmark, maintained on GitHub by creator Georgi Gerganov, an M5 Pro generates 67.59 tokens per second on Llama 2 7B at 4-bit, against 49.64 for an M4 Pro and 35.74 for an 8-core-GPU M6.
The catch: 64GB is the ceiling, and 64GB of Apple memory costs about the same as 128GB of Strix Halo. That's a hard stop for 100B-class models.
Nvidia DGX Spark: CUDA in a box, at a price
The DGX Spark runs Nvidia's GB10 Grace Blackwell chip with a 20-core Arm CPU, 128GB of unified memory at 273GB/s and up to 4TB of storage. Nvidia says one 128GB unit handles models up to 200 billion parameters. Ollama lists GB10 as supported. In the llama.cpp project's DGX Spark benchmark thread, a user posted 45.87 tokens per second for gpt-oss-120b on the stock DGX OS.
Nvidia raised the Founders Edition from $3,999 to $4,699 in February, and its product page no longer lists a price. A 64GB version arrives October 23 through Acer, Asus, Dell, Gigabyte, HP and MSI from $4,999.
The catch: its bandwidth is only a little higher than Strix Halo's, so for plain Ollama chat it costs far more for similar speed. Buy it if you need CUDA, fine-tuning or Nvidia's software stack.
Budget mini PCs: fine for small models
Regular Intel and AMD mini PCs use dual-channel DDR5, about 90GB/s at DDR5-5600. The Beelink SER9 Max with a Ryzen 7 H 255 and 64GB is $1,149, and GMKtec's K12 with the same chip is $359.99 as a barebone you fill with your own RAM and SSD.
These hold big models but run them slowly. They suit 3B to 8B models and always-on home-server jobs. A $1,299 Mac mini M6 has nearly twice the bandwidth for the best free local LLMs at small sizes.
The catch: at about 90GB/s, a 4-bit 27B model crawls along at a few tokens per second, so 64GB here buys capacity, not speed.
What about the Mac Studio and 192GB boxes?
The Mac Studio starts at $2,499 with an M5 Max and 36GB, and its M5 Max reaches up to 614GB/s, double a Mac mini M5 Pro. On the same llama.cpp chart, a 40-core M5 Max hits 119.92 tokens per second. For more memory, the Ryzen AI Max+ PRO 495 brings 192GB: Framework lists it as a pre-order from $6,799, and the GMKtec EVO-X5 Pro is $6,799 with 2TB. Our Mac mini vs mini PC comparison covers the non-AI tradeoffs.
The verdict
The Framework Desktop with the Ryzen AI Max+ 395 and 128GB is the best mini PC for Ollama in 2026. At $3,449 it's the cheapest route to 128GB of GPU-usable memory at 256GB/s, and the GMKtec EVO-X2 is a close second when Framework is out of stock.
Choose the Mac mini M5 Pro with 64GB if you want the fastest, quietest box for models up to about 35B and no driver work. Choose the DGX Spark only for CUDA work. Still choosing an app? Read our Ollama vs LM Studio verdict.
FAQ
What is the best mini PC for Ollama?
The Framework Desktop with AMD's Ryzen AI Max+ 395 and 128GB of memory is the best mini PC for Ollama in 2026. It costs $3,449 for the DIY Edition and its 256GB/s unified memory runs 100B-class models. The GMKtec EVO-X2 128GB at $3,499.99 uses the same chip.
Is a Mac mini good for running local LLMs?
Yes. Ollama runs on the Mac mini's GPU through Metal with no setup, and the M5 Pro model has 307GB/s of memory bandwidth, the fastest of the mini PCs compared here. Its limit is 64GB of memory, which rules out 100B-class models.
How much RAM do I need for Ollama?
16GB runs small 7B to 8B models, 32GB comfortably runs 4-bit 27B to 35B models, and 100B-class models need 96GB to 128GB. On a mini PC, the memory the GPU can use matters more than the CPU.
Is the Nvidia DGX Spark worth it for Ollama?
Only if you need CUDA. It has 128GB of memory at 273GB/s, close to a Ryzen AI Max+ 395 mini PC, but it costs far more after Nvidia's price increases. A 64GB version starts at $4,999 from partners on October 23.
Does Ollama work with the Ryzen AI Max+ 395?
Yes. Ollama lists the Ryzen AI Max+ 395 as supported through ROCm v7 on Linux, and its Vulkan backend works on both Windows and Linux.