Validate UTF-8 leaderboards
Top 20 submissions by each measure. Only submissions that passed the compliance suite are ranked. Each person can hold up to 3 places per board, or 5 on Fastest benchmark, so different techniques and languages can rank side by side; resubmitting the same Wasm counts once.
Size against speed
Gzipped Wasm size against median benchmark time, coloured by the AI models each submitter declared.
Times 1 of these in your browser, one at a time.
| Submission | Server | Your device | Your rank |
|---|---|---|---|
| patrickgwsmith · C · frontier | 2.52 ms | — | — |
Speed, smallest first
Every benchmarked submission in order of gzipped Wasm size, to see whether getting smaller costs speed.
Times 1 of these in your browser, one at a time.
Do the machines agree?
- Server
- Server on the runner corpus
By language
Fastest by language
Each language's fastest submission by median benchmark time, with its gzipped size and whether it uses SIMD.
Smallest download by language
Each language's smallest gzipped Wasm, which sets download time, with its median benchmark time. Open a language to see its next smallest submissions.
Fastest benchmark
Time to render the whole shuffled corpus once after warmup, timed in loops inside Wasm, with input copies subtracted. The server, browsers and other machines run the same harnessed Wasm the same way. Excludes setup and output validation. The corpus is generated (2889 generated inputs, 5.38 MiB in all, up to 255.5 KiB each): labels, prose, escaped HTML, URLs and dense references, with every size band counting for a similar share of the time. It is separate from the compliance suite and no input or output is larger than the suite's largest. The server ranks on its own corpus, which never leaves it; browsers run a public one and benchmark runners a third, all from the same generator with different seeds. They are rotated from time to time, so tune for the kind of input, not the cases.
Speed relative to the leader across every machine, as a geometric mean: each submission is compared with the others on the machines they share, so one measured only on a fast machine isn't flattered by it.
-
1.
patrickgwsmith1.00× 2 machines
- Language
- C SIMD
- Compiler
- Homebrew clang 23.1.5
- AI
- gpt-6.1-sol
Fastest suite
Total time to run the compliance fixture suite.
-
1.
patrickgwsmith5710.81 ms
- Language
- C SIMD
- Compiler
- Homebrew clang 23.1.5
- AI
- gpt-6.1-sol
Smallest Wasm
-
1.
patrickgwsmith1.3 KiB
- Language
- C
- Compiler
- Homebrew clang 23.1.5
- AI
- gpt-6.1-sol