by Inference Hub

The Best Mini PCs for Running LLMs Locally in 2026

Ryzen AI Max mini PCs, NVIDIA's DGX Spark and Apple's Mac mini and Mac Studio compared for local LLM inference: memory, bandwidth, what each can run, and current store prices.

local-inferencehardwaremini-pcgpu

For running large open models at home, the best value right now is a mini PC with AMD’s Ryzen AI Max+ 395 and 128GB of unified memory. The cheapest we found is the GMKtec EVO-X2 at $3,499.99 for 128GB, which holds 120B-class models like gpt-oss-120b and Qwen3.5 122B A10B. If you want more speed per token and can pay more, the Mac Studio with M5 Max has more than twice the memory bandwidth. If you need CUDA, the NVIDIA DGX Spark is the one to get.

Prices below are from each maker’s own store on 22 November 2026, and many are Black Friday sale prices. Memory costs have pushed these machines up a lot this year: 128GB Ryzen AI Max boxes cost about $2,000 at launch and now cost $3,500–4,200.

What matters for local LLMs

Three numbers decide what a machine can do:

  1. Memory the GPU can use. This sets the biggest model you can load. At 4-bit, a model needs about 0.6GB per billion parameters, so a 120B model needs about 70GB. All the machines here use unified memory, which the CPU and GPU share, so a 128GB box can give most of it to the model. A desktop RTX 5090 has only 32GB.
  2. Memory bandwidth. Each generated token reads the model’s active weights once, so bandwidth caps tokens per second. 256 GB/s (Ryzen AI Max+ 395) is fine for mixture-of-experts models and slow for big dense ones; 614 GB/s (M5 Max) is about 2.4 times faster.
  3. Software support. NVIDIA (CUDA) runs everything, including vLLM and most training code. Apple (MLX, llama.cpp) and AMD (ROCm, Vulkan, llama.cpp) are well supported by LM Studio and Ollama for inference, with fewer options for training.

For which models fit in how much memory, and how fast each can go, see Which open models can you run locally?

The machines

MachineChipMemoryBandwidthPrice (store, 22 Nov 2026)
GMKtec EVO-X2Ryzen AI Max+ 39564GB / 128GB256 GB/s$2,199.99 / $3,499.99
Minisforum MS-S1 MaxRyzen AI Max+ 395128GB256 GB/s$3,959 (sale; list $4,949)
Beelink GTR9 ProRyzen AI Max+ 395128GB256 GB/s$4,199 (sale; list $4,699)
GMKtec EVO-X5 ProRyzen AI Max+ PRO 495192GB273 GB/s$6,799
NVIDIA DGX SparkGB10 Grace Blackwell128GB273 GB/s$6,950
Mac mini (M5 Pro)M5 Pro24GB / 64GB307 GB/s$1,699 / $2,699
Mac Studio (M5 Max)M5 Max (40-core GPU)48GB / 128GB614 GB/s$3,099 / $5,099
GMKtec EVO-X1 ProRyzen AI 9 HX 47064GBnot stated$1,599.99

Best value for big models: Ryzen AI Max+ 395 with 128GB

The GMKtec EVO-X2, Minisforum MS-S1 Max and Beelink GTR9 Pro use the same chip and 128GB of LPDDR5X, of which AMD says up to 96GB can be given to the GPU. They perform about the same for LLMs, so choose on price, ports and cooling. The EVO-X2 is the cheapest; the MS-S1 Max is built as a small workstation with more connectivity. The Framework Desktop uses the same chip; its store showed us local-currency prices only, so we haven’t listed it.

They’re a good match for mixture-of-experts models: gpt-oss-120b has a ceiling of roughly 95 tokens a second on this bandwidth, and Qwen3.5 122B A10B about 43. Dense 70B models crawl at a ceiling of about 6.

More memory: Ryzen AI Max+ PRO 495 with 192GB

AMD’s 2026 chip raises the limit to 192GB, with up to 160GB as VRAM, and slightly higher bandwidth (273 GB/s). That’s enough for 230B-class models like MiniMax M2.1 and Qwen3 235B A22B. The GMKtec EVO-X5 Pro is $6,799; Beelink’s 192GB GTR9 Pro is $7,399.

If you need CUDA: NVIDIA DGX Spark

The DGX Spark has the same 128GB and similar bandwidth to the AMD boxes, so generation speed is about the same. It costs about twice as much. What you pay for is CUDA: vLLM, TensorRT-LLM, fine-tuning libraries and most research code run unchanged, and prompt processing is much faster. Partner versions with the same GB10 chip from ASUS, MSI and Dell cost $5,599–8,500 on the NVIDIA Marketplace.

Fastest per token: Mac Studio (M5 Max)

At 614 GB/s, the Mac Studio with M5 Max generates tokens about 2.4 times faster than the 256 GB/s machines on the same model. The 128GB version is $5,099. For the very largest open models, the M5 Ultra goes up to 512GB at 1.2 TB/s, enough for DeepSeek V3.2 at 4-bit; it starts at $5,499 with 96GB, and more memory costs extra.

Smaller budgets

A Mac mini with M5 Pro and 64GB ($2,699) runs the 27–35B models well, like Qwen3.8 27B and Qwen3.6 35B A3B. Ryzen AI 9 HX 370 and HX 470 mini PCs like the GMKtec EVO-X1 Pro ($1,599.99 with 64GB) use ordinary laptop-class memory, so they’re best with small mixture-of-experts models like gpt-oss-20b.

Buy or rent?

A mini PC is a fixed cost with no hourly bill, but renting is often cheaper. An RTX 4090 starts at $0.27 an hour and an RTX 5090 at $0.37 on our GPU price tables. That’s about 18 months of round-the-clock RTX 4090 rental for the price of a 128GB mini PC, and far longer at a few hours a day. The rented cards are faster for models that fit in 24–32GB but can’t hold the 120B class. API prices for open models are lower still: the cheapest gpt-oss-120b API charges $0.092 per million output tokens. We worked through the break-even in our local inference guide.

Buy a local machine for privacy, offline work, no rate limits, or steady all-day use. Rent or use an API if you mostly need bursts.

The mini PCs in this guide

GMKtecfrom $2,199.99

GMKtec EVO-X2

Chip
AMD Ryzen AI Max+ 395
Memory
64GB or 128GB unified (up to 96GB as VRAM)
Bandwidth
256 GB/s
Runs
128GB: gpt-oss-120b, Qwen3.5 122B-A10B, GLM-4.5 Air. 64GB: 30B-class models

128GB / 1TB is $3,499.99. The cheapest 128GB Ryzen AI Max box we found.

Minisforumfrom $3,959

Minisforum MS-S1 Max

Chip
AMD Ryzen AI Max+ 395
Memory
128GB unified (up to 96GB as VRAM)
Bandwidth
256 GB/s
Runs
gpt-oss-120b, Qwen3.5 122B-A10B, Llama 4 Scout, Mistral Small 4

Sale price (list $4,949) for 128GB / 2TB. Built as a small workstation, with more ports than most.

Beelinkfrom $4,199

Beelink GTR9 Pro

Chip
AMD Ryzen AI Max+ 395
Memory
128GB unified (up to 96GB as VRAM)
Bandwidth
256 GB/s
Runs
gpt-oss-120b, Qwen3.5 122B-A10B, GLM-4.5 Air

128GB / 2TB, sale price (list $4,699). A Ryzen AI Max+ PRO 495 version with 192GB starts at $7,399.

GMKtecfrom $6,799

GMKtec EVO-X5 Pro

Chip
AMD Ryzen AI Max+ PRO 495
Memory
192GB unified (up to 160GB as VRAM)
Bandwidth
273 GB/s
Runs
Qwen3 235B-A22B, MiniMax M2.1, plus everything the 128GB boxes run

AMD's 2026 chip. 192GB / 2TB; 4TB is $7,099.

NVIDIAfrom $6,950

NVIDIA DGX Spark

Chip
NVIDIA GB10 Grace Blackwell
Memory
128GB unified
Bandwidth
273 GB/s
Runs
gpt-oss-120b, Qwen3.5 122B-A10B, Llama 4 Scout; CUDA, so most training tools work

128GB / 4TB Founders Edition. Partner GB10 boxes from ASUS, MSI and Dell cost $5,599 to $8,500.

Applefrom $3,099

Mac Studio (M5 Max)

Chip
Apple M5 Max
Memory
48GB or 128GB unified
Bandwidth
614 GB/s (40-core GPU)
Runs
128GB: gpt-oss-120b, Qwen3.5 122B-A10B, at more than twice the bandwidth of the AMD boxes

128GB is $5,099. The M5 Ultra (96GB to 512GB, 1.2 TB/s) starts at $5,499.

Applefrom $1,699

Mac mini (M5 Pro)

Chip
Apple M5 Pro
Memory
24GB or 64GB unified
Bandwidth
307 GB/s
Runs
64GB: Qwen3.8 27B, Qwen3.6 35B-A3B, gpt-oss-20b, Gemma 4 26B

The 64GB model is $2,699; the 24GB base only fits small models.

GMKtecfrom $1,599.99

GMKtec EVO-X1 Pro

Chip
AMD Ryzen AI 9 HX 470
Memory
64GB
Runs
gpt-oss-20b, Qwen3.6 35B-A3B and other small MoE models

A budget option: ordinary laptop-class memory, so dense 27B+ models are slow.

Starting prices from each maker's own store on November 22, 2026; they change often, so check the store. "Runs" is a rough guide for 4-bit quantized models.

Subscribe to our newsletter

News and updates, 1–2 emails per month. No spam.