Repository navigation
First public release for the 2026 Arm AI Optimization Challenge.
Native aarch64 evidence:
- 74.55% smaller model (86.22 MB to 21.94 MB)
- 59.47% lower mean latency (19.31 ms to 7.83 ms)
- 146.74% higher throughput
- 58.58% lower process RSS delta
- 0.98344814 full-output cosine similarity (quality gate passed)
Includes the quantizer, isolated benchmark workers, reusable composite action, public dashboard, raw JSON and Markdown reports, methodology, and tests.