Local AI quality and performance
This page defines a test protocol. Measurements and downloadable result artifacts have not been published.
Full benchmark protocol
- Generate synthetic small, medium, large, malformed and boundary-case inputs with known properties.
- Record engine version, browser, operating system, memory and processor. Use the same device and settings for each comparison.
- Run a warm-up, then at least five measured repetitions. Report median and spread alongside failures.
- Check outputs independently for preserved content and declared losses. Keep the original and generated artifacts available for reproduction.
- Test cancellation, offline reload, repeated-job memory usage and a network-marker input.
Results table
Pending. No zero-time, zero-upload, compression-ratio or quality score is invented here.