Privacy

Web AI Benchmark is a tool that benchmarks on-device AI inference inside your browser. This page describes what data the app collects, where it goes, and how to remove it.

What we never collect §

• Model weights, prompts, and generated text never leave your browser. Benchmarks run locally via WebGPU, WebNN, or WebAssembly. The runtime fetches model files directly from HuggingFace into your browser; we don't proxy or copy them.
• No analytics, ads, or behavioral tracking. No third-party SDKs.

What we collect when you sign in §

If you sign in with GitHub or Google, the OAuth provider sends us:

• Your email address.
• Your provider profile (display name, avatar URL).
• A user identifier issued by the provider.

We use this to scope your benchmark history and recipes to your account. You can update your display name and upload a custom avatar from Account.

What we collect when you click "Upload results" §

The "Upload results" toggle on the run pages is opt-in and only available while signed in. When enabled, after each benchmark we save:

• Benchmark metrics (latency, throughput, TTFT, TPS, decode time, etc.).
• Run configuration (backend, dtype, framework version, iterations, prompt-token count, max-new-tokens).
• The HuggingFace model identifier and file path you ran — never the model bytes themselves.
• Hardware details you typed in: CPU, OS, GPU/NPU driver versions.
• Auto-detected environment: browser name and version, GPU adapter string.
• A run timestamp and the user identifier from your account.

None of this is collected when "Upload results" is off, even when you're signed in.

Where saved data lives §

Account info, profiles, recipes, and saved benchmark results are stored in a hosted database. Access is gated by row-level security so users can only read and delete their own rows. Saved benchmark results themselves are immutable; to update a measurement, run the benchmark again.

Data that stays only in your browser §

• Model file cache (OPFS): downloaded model weights so subsequent runs don't re-download. Wipe via the "Clear Model Cache" button in the footer or your browser's site-data tools.
• Model catalog (IndexedDB): a cached copy of the public model catalog used by the Browse page.
• Preferences (localStorage): table column choices, filter values, recently-used hardware fields.
• Interrupted-run state (localStorage): so the app can offer to resume a run after a crash or accidental close.

None of these are sent to our servers.

Third parties §

• HuggingFace — model files are fetched from huggingface.co directly to your browser. They see your IP address and which models you fetch, subject to their own privacy policy.
• GitHub OAuth / Google OAuth — only when you sign in.

Retention and deletion §

Saved benchmark results are kept until you delete them or close your account. We don't auto-expire or anonymize.

• Delete individual results from /inference/results or /llm/results.
• Delete recipes from /inference/recipe or /llm/recipe.
• To delete your account and all associated data, email Show email.

Site administrators may also delete any user data (accounts, results, recipes, or shared links) at their discretion — for example to remove spam, content that violates the terms of service, or at a user's request.

Your rights §

Depending on where you live, you may have rights to access, correct, export, or delete the personal data we hold about you. We honour these requests for any signed-in user. Contact us at Show email.

Children §

The service is not directed at children under 18 (or the equivalent minimum age in your jurisdiction). We don't knowingly collect data from them. If you believe a child has signed up, contact us and we'll remove the account.

Changes §

We'll update the "Last updated" date when we change this page. For material changes that affect how saved data is used, we'll notify signed-in users via the app.

Contact §

For privacy questions or data-deletion requests: Show email.