Benable — create shareable lists of things you recommend!
E.g., products you love, local businesses, travel recs - you can add anything to a Benable list!

Self-Hosted AI Agents: Best Hardware to Run Local LLMs (2026)

Purple Star emoji 12 items
I'm a backend engineer who builds AI agents every day, and I run the Ship With AI YouTube channel. These are the best devices to self-host AI agents and local LLMs in 2026: no cloud bills, no data leaving your home. From a cheap Raspberry Pi to 128GB...
Sections
5
 
 
 
 
 

128GB AI mini supercomputers

Desk-size boxes with 128GB of shared memory. They run 70B to 120B models at home.

 
Darjus profile picture
The easiest way to get 128GB of fast memory with full NVIDIA CUDA. Runs 120B models like gpt-oss-120b at home, and you can link two for bigger ones. Pricey, but it just works.
NVIDIA DGX Spark (128GB)
 
Darjus profile picture
Same NVIDIA GB10 chip and 128GB memory as the DGX Spark, but usually cheaper. My pick if you want Spark power for less money. ASUS also sells a cable to stack two.
ASUS Ascent GX10 (128GB)
 
Darjus profile picture
Best value 128GB box. The AMD Ryzen AI Max+ 395 runs 70B models in Ollama or LM Studio, and it works with Windows or Linux. A great cheaper alternative to the GB10 boxes.
GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB)

Macs for local LLMs

Apple Silicon shares fast memory between CPU and GPU, so Macs run big models quietly. Memory size matters most.

 
Darjus profile picture
Best value Mac for local AI. Order it with 64GB: it runs 32B models fast and even 70B slowly. Tiny, silent and uses little power, so it's perfect as a 24/7 agent server.
Apple Mac mini M5 Pro
 
Darjus profile picture
The best all-round Mac for AI. With 128GB it runs gpt-oss-120b at about 88 tokens per second, and it stays quiet on your desk. Need even bigger models? The M5 Ultra goes up to 256GB.
Apple Mac Studio M5 Max

AI laptops

Run local models anywhere, even with no internet.

 
Darjus profile picture
Best laptop for local LLMs. Same M5 Max chip as the Mac Studio, so you can run 120B models on a plane with no internet. Get 128GB if your budget allows.
Apple MacBook Pro 14" M5 Max
 
Darjus profile picture
A Windows option with the same AMD AI Max+ 395 chip as the mini PCs. The 64GB version runs 30B models on the go, and it's a 2-in-1 tablet that still plays games.
ASUS ROG Flow Z13 (Ryzen AI Max+ 395, 64GB)

AI NAS & home servers

Keep your data, Docker apps and agent tools at home, running 24/7.

 
Darjus profile picture
An AI NAS with a built-in local assistant (Uliya) that searches and answers questions about your own files. 64GB RAM and 6 bays, so your photos and docs never leave home.
UGREEN NASync iDX6011 Pro AI NAS
 
Darjus profile picture
Cheaper 4-bay NAS that runs Docker well. Great home for n8n, Open WebUI and a vector database for your AI agents, while a bigger box does the heavy model work. 10GbE is a nice bonus.
UGREEN NASync DXP4800 Plus

Budget & always-on AI boxes

Small, cheap and low power. Great for always-on agents, home automation and learning.

 
Darjus profile picture
The cheapest always-on agent server. Perfect for Home Assistant, n8n or a small bot that calls the Claude or OpenAI API. 16GB is enough for small local models too.
Raspberry Pi 5 (16GB)
 
Darjus profile picture
Turns a Pi 5 into a tiny local AI box. The Hailo-10H chip adds 40 TOPS and its own 8GB RAM, so it runs small LLMs and vision models without slowing the Pi down. Very low power.
Raspberry Pi AI HAT+ 2 (40 TOPS)
 
Darjus profile picture
NVIDIA's small AI dev kit with real CUDA. Great for learning, robots and camera-based agents. The 8GB is tight for LLMs, so stick to small models.
NVIDIA Jetson Orin Nano Super Developer Kit