We would love to accept your patches and contributions to this project.
Contributions to this project must be accompanied by a Contributor License Agreement (CLA). You (or your employer) retain the copyright to your contribution; this simply gives us permission to use and redistribute your contributions as part of the project.
If you or your current employer have already signed the Google CLA (even if it was for a different project), you probably don't need to do it again.
Visit https://cla.developers.google.com/ to see your current agreements or to sign a new one.
This project follows Google's Open Source Community Guidelines.
If you believe a model should be included in our benchmark, please submit a pull request to add it as a new workload.
As of Q1 2026, the benchmark is under active development, and we are refining aspects such as the scoring calculation. We welcome your input on design considerations or other general feedback. Please open an issue to start a discussion.
To propose a new workload, please add it to the resources/experimental directory and submit a pull request. In your pull request, please explain the workload's relevance and provide technical details about its integration.
- Add any required dependencies in
resources/experimental/package.json. - Create a subdirectory
resources/experimental/src/<workload>/containing your benchmark entry point(s) (e.g.benchmark.mjs), similar to the existing workloads. If the workload compiles C/C++ or Rust to WebAssembly, include abuild-wasm.shscript, check in the prebuilt output underresources/experimental/src/<workload>/build/, and register the workload inresources/experimental/build-wasm-workloads.sh(seeresources/experimental/README.md). - Ensure that models are self-contained and not loaded remotely from a CDN or the Hugging Face Hub at runtime. Place model files under
resources/experimental/models/<workload>/and track large binary files with Git LFS in.gitattributes. - Add licensing information for the library and/or model in
resources/experimental/licenses/NOTICE.mdand add any required license file inresources/experimental/licenses/. - Add an entry and an
HtmlWebpackPlugin(usingsrc/console-runner.htmlas the template) for the new workload inresources/experimental/webpack.common.js. - Update
.gitignoreor.gitattributesif needed. - Run
npm run buildin the repository root directory. Alternatively, you can runnpm installandnpm run buildinsideresources/experimental. - Add the workload to
resources/default-tests.mjswith (optionally) theExperimental-prefix andexperimentaltag, analogous to the existing workloads. - Serve the overall runner via
npm run devin the repository root directory. - The experimental workloads are not part of default set of workloads. Browse to
http://localhost:8080/?developerMode, select the added workloads alongside any other workloads and click "Start Test" to run the benchmark. You can find more instructions about how to use the Developer Mode in Developer Mode and Custom Parameters. - To run individual experimental workloads, navigate to
http://localhost:8080/resources/experimental/dist/<workload>.html.
All submissions, including submissions by project members, require review. We
use GitHub pull requests
for this purpose. Run npm run fix to format and lint your changes before
submitting a pull request.