About This Site
Arao AI is a benchmark database for local AI deployment: it answers "can my machine run this model, with what setup, and how fast". This page covers who we are, what is covered, and when the numbers might mislead you.
What this site is
This site is for people preparing to run models on their own hardware — with a GPU or a machine at hand, you can check whether a given setup runs a model and roughly how fast, before spending money. Online, these questions either go unanswered or only get vendor peak specs, and peak specs do not predict real throughput.
Our approach: collect public benchmarks scattered across the web, normalize them into comparable units, label evidence levels, and present data gaps honestly. The site's principle is "no record ≠ can't run": if a combination isn't in the database, it means nobody has published a test — not that it won't work. We don't fill gaps with speculation, and we never treat absence as a negative conclusion.
Known limitations
The variables below materially affect results but are not modeled in this database yet. Take them into account when making decisions from these numbers; if your scenario falls into one of them, treat the numbers as order-of-magnitude references only.
- Context length: a record usually reflects one context-length setting. Longer contexts squeeze VRAM with KV cache and noticeably reduce throughput; the database does not annotate this separately yet.
- Concurrency: most public benchmarks are single-request results. Total throughput under concurrent load is not the same as single-stream Decode.
- KV cache precision and batching parameters: these framework-side settings affect VRAM and speed, and public data often omits them.
- Driver, framework, and quantization-tool versions: the same hardware can produce different results under different version combinations; records are labeled, but cross-version comparability is not guaranteed.
- Individual hardware variation and cooling: laptops and prebuilt machines have different power and thermal limits, and the same chip name can deliver different numbers.
These limitations are real, not legal boilerplate: they determine how far a number can be extrapolated. As coverage grows, the most impactful one (context length) will be modeled first.
Corrections and contact
Every record is curated by hand, so mistakes are possible. If you spot one, please write in — the most helpful reports include the original source: a link, a configuration note, anything we can verify against and then fix.
If you happen to have benchmark results we haven't collected, those are welcome too; see "Data notes" for inclusion criteria.
Sources and compliance
Benchmark results on this site come from publicly published material, each labeled with its source type and link. We cite data and conclusions only, never republish the full original text; excerpts appear in minimal form for verification, and copyright stays with the original authors.
This site has no affiliation with and receives no sponsorship from any hardware, model, or framework vendor mentioned; no payment of any kind is accepted for including a record. Missing data and negative results are shown all the same — the other side of "no record ≠ can't run": let the data speak for itself.
If you are the original author of a data point and want the citation adjusted or the record taken down, write to us via the Contact page.