TrialMirror: Local AI hardware test bench

Run real models on your hardware before buying that local-AI machine—see speed and memory first.

Problem

Vendors keep pushing expensive local-AI computers, but buyers don't know how fast their models will actually run or how much memory they need. Existing benchmarks are either data-center oriented or ignore edge inference.

Solution

A desktop app with standardized on-device LLM inference tests. Users download specified open models or bring their own, run one-click benchmarks for tok/s, first-token latency, peak memory, and power estimates, then generate comparable hardware reports.

Users

Developers, researchers, and procurement teams considering local AI hardware like Mac Studio, AI PCs, or external accelerators.

MVP

3 standard model test suites; one-click benchmark recording tok/s and memory; shareable result cards; community-submitted hardware comparison.

Revenue

Free benchmarking; Pro subscription for more models and detailed power analysis; sponsored leaderboard from hardware vendors.

Why now

Local AI hardware marketing is hot but neutral edge benchmarks are scarce, creating strong decision anxiety. An indie can ship a lightweight benchmark and grow via community data.