Benchmark 01 is in progress. Leave your email to get the results first

Your own local AI. On your desk. Not in their cloud.

Independent benchmarks of compact AI computers — from mini PCs to AI workstations — for power users and small businesses who want AI that stays private and has no limits.

See the test classes Read the methodology
Ryzen AI Max+ PRO 495Ryzen AI Max+ 395Ryzen AI 9 HX 470Core Ultra X9Apple M5 ProApple M6

01 / why local

Two reasons people move AI off the cloud.

Confidential by design.

  • Prompts, documents and answers never leave your machine.
  • Nothing logged, retained or used for training by a third party.
  • Client files, HR data, source code stay under your control — GDPR and professional secrecy made simpler.
  • Runs fully offline, even on an isolated network.

No quotas. No limits.

  • No message caps, no rate limits, no “try again in 3 hours”.
  • No per-seat or per-token billing: one hardware cost, used as much as you want.
  • Long documents and overnight batch jobs run as long as they need.
  • The model never changes behind your back.
Plus full controlPick open-weight models, quantization, context size and runtime.

02 / who it's for

Your models, your data, your hardware.
The only question is which hardware.

Power users

  • Run 70B dense and 120B MoE models on your desk
  • Coding assistants and agents that stay on your network
  • Plug into your own tools: home automation, workflows, scripts
  • Tune quantization, context size and runtime

Small businesses

  • Ask questions across contracts, invoices and procedures
  • Transcribe and summarise meetings on site
  • One machine serving the whole team
  • A predictable cost instead of per-seat AI subscriptions

03 / test classes

Three classes.
Compared like for like.

Machines are ranked within their class — never a 1,500 € box against a 7,000 € workstation. Value is shown as speed per euro.

CLASS 1 · ESSENTIAL

Ryzen AI 9 HX 470

~0

GB/s max

0

GB RAM max

memory bandwidthvs. 273 GB/s

Everyday assistant, 8–30B models, NPU workloads.

CLASS 2 · PERFORMANCE

Ryzen AI Max+ 395

0

GB/s

0

GB unified

memory bandwidthvs. 273 GB/s

70B dense, 120B MoE — the local-LLM sweet spot.

CLASS 3 · AI WORKSTATION

Ryzen AI Max+ PRO 495

0

GB/s

0

GB unified

memory bandwidthvs. 273 GB/s

The largest open models, long context, several models at once.

ReferenceApple Mac mini M6 · 32 GB · 170 GB/sApple Mac mini M5 Pro · 64 GBPeak figures from manufacturers. Measured values per machine.

04 / methodology

Same models. Same versions.
Weeks on the bench, not minutes.

speedPrompt processing and generation, 8B to 120B-class models, short to very long context.
enduranceMulti-hour runs: does throughput hold? Temperatures, fan noise, throttling over time.
efficiencyPower at the wall, tokens per watt, and what the NPU really adds.
daily useEach machine is lived with day to day: chat, documents, code, transcription.
setupWindows and Linux, GPU memory allocation, drivers and ROCm maturity.
No paid rankingsPlacement can't be bought.
Supplied hardware disclosedEvery review says who provided the machine.
Method firstPublished before the first result.

tests in progress

Benchmark 01 is underway.

Leave your email and you'll be the first to know when the results are published. One email, nothing else.

Your email is only used to send you the benchmark results. Unsubscribe anytime. See privacy.