Infino benchmarks
Every number on this page links to the run behind it and a way to reproduce it. Where another engine is faster, the page says so.
SQL analytics: ClickBench
On the public ClickBench board, Infino ranks in the top 11% for SQL query speed on the leaderboard machine (c8g.metal-48xl, hot runs), out of roughly 501 systems.
It stores the same 100M-row dataset in the top 6% for size on c6a.4xlarge, and on the leaderboard machine stays ahead of every ClickHouse-on-Parquet and DataFusion build.
Filtered to the search engines on the board, Infino ranks #1 for SQL analytics (combined metric), ahead of Quickwit, Elasticsearch, SigLens and ParadeDB.
| machine | SQL query speed | storage size |
|---|---|---|
| c8g.metal-48xlleaderboard machine, 192 vCPU Graviton4 | top 11% | top 7% |
| c7a.metal-48xlhigh-core machine, 192 vCPU AMD | top 15% | top 7% |
| c6a.4xlargereference machine, 16 vCPU AMD | top 40% | top 6% |
100M rows, 43 queries, hot runs (the best of the repeated runs), lower is faster. Percentile is Infino's position among the systems on the live board for that machine (~501 systems), snapshot 2026-10-06; it shifts as systems are added, so compare machine first. Sources: the official board · Infino's directory in ClickBench, which is how to reproduce it.
Search latency on object storage
Measured by the engine's own harness in CI, on 1M-document tables in Azure Blob with a warm cache.
| query | shape | result |
|---|---|---|
| full-text (BM25) | single rare term, top-10, warm + fetch | 166 µs p50 |
| vector | top-10 cosine, 1024 dimensions, post-drain | 663 µs p50 · 0.992 recall@10 |
| hybrid | hybrid_search through SQL, 10 rows | 3.47 ms p50 |
| ingest | full-text / vector / SQL tables | 62.1 / 7.8 / 17.0 K docs/s |
- storage
- Azure Blob, 1M-document tables, warm cache
- hardware
- Intel Xeon 6973P-C, 4 cores pinned, 31 GiB
- engine
- commit bd7caa3, measured in CI
- harness
- cargo bench in the engine repository
Reproduce with cargo bench in the engine repository. Every cell and its flags are in benches/README.md, next to the CI run and raw logs these numbers come from. First-touch cost on object storage is covered in tradeoffs.
Vector databases: VectorDBBench
Infino runs through VectorDBBench, the open vector database harness, the same way as any engine on its public board. The open viewer shows every engine measured on one VM size at Cohere 1M and 10M, each row linked to its result file.
open the VectorDBBench results →
git clone https://github.com/infino-ai/VectorDBBench && cd VectorDBBench pip install -e . init_bench
Head-to-head with libraries
retrievalbench runs Infino against other engines on the same data, in the open, with the harness and every result file public.
- Embedded vector libraries. turbovec and FAISS in one process, same queries, brute-force ground truth, each library through its own public API.
- Full-text at Wikipedia scale. Tantivy, Lucene and others through Search Benchmark, the Game, single-threaded over full Wikipedia.
Results, charts and reproduce commands: retrievalbench. Published results are CC BY 4.0, so they can be quoted in a comparison that disagrees with ours. If you maintain an engine measured there and think it is configured unfairly, open an issue.
Common questions
Why is hot the number to compare on ClickBench?
ClickBench runs each query three times. Cold is the first run and hot is the best of the other two. Infino took its cold runs with the OS page cache dropped before every query, while the reference engines run cold against a warm OS cache, so the cold columns measure different things. Hot is measured the same way for every engine.
Which engines are ahead of Infino on ClickBench?
By hot query speed on the leaderboard machine, Infino sits in the top 11% of the field. The systems ahead are native-format or in-memory engines and DuckDB and ClickHouse builds, led by Umbra, CedarDB and DuckDB. Infino stays ahead of every DataFusion build and every ClickHouse-on-Parquet variant on that machine. The full field is the live board.
What does a query cost the first time it touches object storage?
The first query to touch a file pays object-store round trips, hundreds of milliseconds, before its bytes are cached. Warm and repeated queries do not pay it. The figures above are warm, and the tradeoffs page in the docs states the cold envelope.
Can I reproduce these numbers?
Yes. ClickBench runs from the infino directory in the upstream ClickBench repository. The search latencies run with cargo bench in the engine repository. VectorDBBench and the library head-to-heads run from retrievalbench, and each published result records the engine commit and the command.
Measured. Now try it on your own data.
pip install infinoquickstart →