The local AI hardware guide

Choose a machine.
See what actually runs.

Real computers. Real creator tests. Find hardware for your work, inspect the exact configuration, and watch the evidence before you buy.

Explore 30 hardware optionsHow we choose hardware
01 / Memory

Start with the model

Model size, quantization and context all change memory needs. A GPU’s VRAM and a Mac’s shared memory describe different designs.

02 / Workflow

Match the work

Private chat, coding, image generation and robotics need different software and hardware. An NPU alone does not guarantee model support.

03 / Ownership

Check the whole cost

A GPU needs a suitable host PC and power supply. A developer kit may need storage, cooling and setup. Check local warranty and current pricing.

Find your starting point

Hardware, with the evidence attached

Older systems remain useful budget choices. Newer hardware is included when a real local-model test is available.

30 researched options · Select up to three to compare

Compare (0/3)
Apple iMac M3 (2023): Bijan Bowen / Ominous Industries's hands-on review
Apple Silicon · All-in-one desktop

Apple iMac M3 (2023)

Reviewed unit: M3 iMac with 16 GB memory; port/GPU tier and SSD unspecified.

Chat & learning

Tested models
Llama 3 8B

Review by Bijan Bowen / Ominous Industries

See tests & buying guide
Apple Mac mini M4 (2024): Alex Ziskind's hands-on review
Apple Silicon · Mini desktop

Apple Mac mini M4 (2024)

Reviewed unit: base M4 Mac mini; video source identifies base model but retrieved material does not explicitly confirm SSD/RAM.

Home servers

Tested models
Llama 3.2 1B

Review by Alex Ziskind

See tests & buying guide
Apple Mac mini M4 Pro (2024): Andrej Baranovskij / Katana ML's hands-on review
Apple Silicon · Mini desktop

Apple Mac mini M4 Pro (2024)

Creator identifies Mac mini M4 Pro; reviewed memory, CPU/GPU bin and SSD not disclosed in the accompanying writeup.

Chat & learningCoding & agents

Tested models
Gemma 4 31B

Review by Andrej Baranovskij / Katana ML

See tests & buying guide
Apple Mac Studio M2 Ultra (2023): NetworkChuck's hands-on review
Apple Silicon · Desktop workstation

Apple Mac Studio M2 Ultra (2023)

Reviewed setup: five M2 Ultra Mac Studios,64 GB memory each. Later section tests one 64 GB Studio using Ollama.

Large-model experimentsImage & video generation

Tested models
Llama 3.3 70B

Review by NetworkChuck

See tests & buying guide
Apple Mac Studio M3 Ultra (2025): NetworkChuck's hands-on review
Apple Silicon · Desktop workstation

Apple Mac Studio M3 Ultra (2025)

Reviewed unit: M3 Ultra 32-core CPU,80-core GPU,512 GB unified memory; macOS 26.5.1. Historical 512 GB configuration is confirmed by reviewer, while current archived Apple page lists 96/256 GB.

Large-model experimentsDocuments & automation

Tested models
Qwen3.8 27B · Qwen3.5 122B-A10B

Review by NetworkChuck

See tests & buying guide
Apple Mac Studio M5 Ultra (2026): NetworkChuck's hands-on review
Apple Silicon · Desktop workstation

Apple Mac Studio M5 Ultra (2026)

Reviewed unit:36-core CPU,80-core GPU,256 GB memory; macOS 27.0; Apple review loan.

Large-model experimentsImage & video generation

Tested models
Qwen3.8 27B · Gemma 4 31B

Review by NetworkChuck

See tests & buying guide
Framework Desktop (Ryzen AI Max+ 395): Donato Capitella's hands-on review
Mini PCs · 4.5L desktop

Framework Desktop (Ryzen AI Max+ 395)

Pre-production Framework Desktop, Max+395; Linux/Fedora ROCm 7 toolbox. Capacity not explicitly stated in accessible video description.

Chat & learningImage & video generation

Tested models
Qwen-Image · Wan 2.2

Review by Donato Capitella

See tests & buying guide
Beelink GTR9 Pro: ServeTheHome / Patrick Kennedy's hands-on review
Mini PCs · Mini PC

Beelink GTR9 Pro

128 GB; 96 GB GPU /32 GB CPU memory allocation; Windows LM Studio.

Chat & learningHome servers

Tested models
gpt-oss-120b · Llama 3.3 70B

Review by ServeTheHome / Patrick Kennedy

See tests & buying guide
GMKtec EVO-X2: Country Boy Computers's hands-on review
Mini PCs · Mini PC

GMKtec EVO-X2

128 GB, Windows 11, llama.cpp; compared CPU and iGPU execution with EVO-T1.

Chat & learning

Tested models
gpt-oss-20b · Qwen3.5 35B · Gemma 4 26B

Review by Country Boy Computers

See tests & buying guide
GMKtec EVO-T1: Country Boy Computers's hands-on review
Mini PCs · Mini PC

GMKtec EVO-T1

96 GB DDR5, Windows 11, llama.cpp.

Chat & learning

Tested models
gpt-oss-20b · Qwen3.5 35B · Gemma 4 26B

Review by Country Boy Computers

See tests & buying guide
Minisforum MS-S1 Max: ServeTheHome / Ryan Smith and Patrick Kennedy's hands-on review
Mini PCs · 3.5L mini workstation

Minisforum MS-S1 Max

128 GB/2 TB Windows 11Pro, Performance mode.

Chat & learningLarge-model experimentsHome servers

Tested models
Qwen3-Next 80B

Review by ServeTheHome / Ryan Smith and Patrick Kennedy

See tests & buying guide
HP Z2 Mini G1a: Donato Capitella's hands-on review
Mini PCs · Mini workstation

HP Z2 Mini G1a

HP Z2 Mini G1a,128 GB,Linux,Fedora toolbox; tuned kernel/memory allocation.

Chat & learningLarge-model experiments

Tested models
GLM4.5-Air106B · Qwen3-235B

Review by Donato Capitella

See tests & buying guide
ASUS Ascent GX10: Techno Tim's hands-on review
AI workstations · Desktop AI appliance

ASUS Ascent GX10

Single128 GB GX10 with vLLM for first test; second test used two128 GB GX10s connected over QSFP.

Coding & agentsHome servers

Tested models
Qwen3.6-35B-A3B · MiniMax M2.7

Review by Techno Tim

See tests & buying guide
Dell Pro Max with GB10: ServeTheHome / Patrick Kennedy's hands-on review
AI workstations · Desktop AI appliance

Dell Pro Max with GB10

Two Dell-supplied128 GB systems for independent workflow/model tests; do not assume models were split across both.

Documents & automation

Tested models
gpt-oss-120b · gpt-oss-20b

Review by ServeTheHome / Patrick Kennedy

See tests & buying guide
Bosgame M5: ServeTheHome / Patrick Kennedy's hands-on review
Mini PCs · Mini PC

Bosgame M5

128 GB M5; reviewer compared performance modes and Qwen inference across StrixHalo systems.

Chat & learning

Tested models
Qwen3.6-35B · Qwen3.8-27B

Review by ServeTheHome / Patrick Kennedy

See tests & buying guide
NVIDIA DGX Spark 128 GB: Alex Finn's hands-on review
AI workstations · Compact complete Arm Linux computer

NVIDIA DGX Spark 128 GB

Single DGX Spark; exact SSD and software build not established in indexed video description

Coding & agents

Tested models
Qwen 3.6 27B

Review by Alex Finn

See tests & buying guide
NVIDIA Jetson Orin Nano Super Developer Kit: Bijan Bowen's hands-on review
Edge AI developer kit · Board computer requiring user-provided storage

NVIDIA Jetson Orin Nano Super Developer Kit

Orin Nano Super running Jetson containers, Ollama, LlamaIndex and Streamlit; storage capacity not confirmed

Documents & automationEdge & robotics

Tested models
Llama 3.2 (size not specified in video description)

Review by Bijan Bowen

See tests & buying guide
NVIDIA Jetson AGX Thor Developer Kit: STH Labs / ServeTheHome's hands-on review
Edge AI developer kit · Complete developer platform

NVIDIA Jetson AGX Thor Developer Kit

T5000 developer kit with 128 GB and included 1 TB SSD; STH describes short test window using workloads close to NVIDIA reference material

Coding & agentsEdge & robotics

Tested models
Llama 3.1 8B

Review by STH Labs / ServeTheHome

See tests & buying guide
NVIDIA GeForce RTX 4090 24 GB: GamersNexus / Steve Burke's hands-on review
GPU upgrade · PCIe GPU card; not a complete computer

NVIDIA GeForce RTX 4090 24 GB

GamersNexus comparison bench; exact LLM host CPU/RAM not confirmed. See the original test method.

Chat & learningLarge-model experiments

Tested models
DeepSeek Llama 8B Distil (review label)

Review by GamersNexus / Steve Burke

See tests & buying guide
NVIDIA GeForce RTX 5090 32 GB: GamersNexus / Steve Burke's hands-on review
GPU upgrade · PCIe GPU card; not a complete computer

NVIDIA GeForce RTX 5090 32 GB

GamersNexus comparison bench; exact LLM host CPU/RAM not confirmed. See the original test method.

Chat & learningLarge-model experiments

Tested models
DeepSeek Llama 8B Distil (review label)

Review by GamersNexus / Steve Burke

See tests & buying guide
AMD Radeon RX 6700 XT 12 GB: Jeff Geerling's hands-on review
GPU upgrade · PCIe GPU card; not a complete computer

AMD Radeon RX 6700 XT 12 GB

Raspberry Pi 5 plus external RX 6700 XT; patched amdgpu kernel and llama.cpp Vulkan; not an ordinary plug-and-play desktop configuration

Chat & learning

Tested models
Llama 3.2 3B Instruct

Review by Jeff Geerling

See tests & buying guide
AMD Radeon RX 7900 XT 20 GB: Jeff Geerling's hands-on review
GPU upgrade · PCIe GPU card; not a complete computer

AMD Radeon RX 7900 XT 20 GB

Raspberry Pi 500+ 16 GB using M.2-to-OCuLink dock, external PSU, custom Linux kernel patch and Vulkan. Benchmark is specific to this unusual host.

Chat & learning

Tested models
Llama 3.2 3B Instruct · Qwen2.5 14B Instruct

Review by Jeff Geerling

See tests & buying guide
Side by side

Compare the reviewed machines

These are different tests, not a shared benchmark. Use the review methods before comparing speed.

Select up to three cards above to build your comparison.

Our research standard

A review is evidence. A specification is a boundary.

Manufacturer specifications

We link the official product or technical specification page. The reviewed unit is separate from the maximum configuration offered for sale.

Hands-on creator sources

Each profile links a YouTube review and its supporting source. We record named models, engine and quantization only when the source identifies them. We do not claim to have independently benchmarked these machines.

Limits stay visible

Unknown settings, software gaps and setup difficulty matter. Results from different reviews cannot be compared as a controlled speed ranking. Our buying advice is an editorial interpretation of the cited evidence.

Community experience

Owner reports add context when available, but are not lab benchmarks. Share your exact machine, model file, engine, context and outcome so others can reproduce your result.

Share a hardware test →