Measuring the future of sovereign AI.
Hqllr is built for speed and verifiability. This repository showcases benchmark results across devices, corpora, and cluster scales.
- Rapid Ingestion — millions of words processed per second per core.
- Atomisation — sentence- and paragraph-level indexing with zero duplication.
- Scalable — parallel ingestion scales linearly with cores.
- Efficient — optimised storage with filesystem compression.
- Portable — consistent performance across mobile, desktop, and clusters.
- 📚 12,600 books (Project Gutenberg corpus) ingested in 4.5 minutes on an 8-core tablet (using 4 cores).
- ⚡ Throughput doubled in the revised engine (Sep 2025), hitting new record speeds per core.
- 📊 Consistent scaling: near-linear atoms/sec increase with core count.
- Benchmarks run on reference hardware (documented per test).
- Tests are repeatable, using public corpora (e.g., Project Gutenberg).
- Results reported in:
- Files/sec
- Atoms/sec (sentence, paragraph, page modes)
- Scaling per core
- Memory footprint (RSS)
- Regular milestone benchmarks
- Comparative scaling (mobile vs. desktop vs. cluster)
- Public benchmark harness (later release)
- Community benchmark submissions
Performance is not just speed — it’s sovereignty:
- Fast enough to run locally, even on constrained devices.
- Efficient enough to scale to enterprise clusters.
- Transparent enough to verify.