runtime backend demo

What can this device run, and does a small model answer a prompt on it? Everything happens in this tab. Nothing is uploaded, and nothing is requested from this site's models/ folder until you click the download button.

?. The Pages build is .github/workflows/pages.yml. Asset manager demo.

Sprint 5 (18 Nov 2026): Working: runtime beta 0.3.0-beta.0 on spec release 0.5.0: capability profile, signed-manifest verification (now with an expected-id check that refuses a validly signed manifest of another product), download, hash check and quota eviction of synthetic assets, resume, and a Tamper lab of fourteen cases that shows the exact refusal reason, with 0 requests after a manifest refusal. The test-model page adds embeddings through the worker host, in-worker hashing of every model file, a token budget and Stop. This page's policy is now connect-src 'self': it can only reach its own origin. Not yet: WebLLM (cut), Windows and 8 GB measurements (not measured on 8 GB), key revocation, a real signing key.

What this demo shows

Same prompt answered on WebGPU and on WebAssembly with the test model SmolLM2-135M-Instruct (q4), in Chrome on an Apple-silicon Mac only. This is not the Hindi production model; Windows and the 8 GB no-GPU laptop are not measured.

New in Sprint 3: the engine runs in a dedicated worker, so the page stays responsive while WebAssembly generates, and the library ends an aborted answer in about 250 ms (evidence/sprint-3/D1/worker/); this page requires Trusted Types (one policy); and the load timing and memory read-out below shows what the Airplane Mode DevTools panel shows.

Planned production model: Qwen3-1.7B (Apache-2.0), WebGPU q4f16 1.43 GB, WASM q8 1.74 GB. Provisional until the Sprint 5 measurement on an 8 GB laptop; contingency Qwen3-0.6B. Loaded once in a browser: a local, unscored WebGPU smoke run of five synthetic Hindi prompts on an Apple-silicon Mac, which needed ONNX Runtime graph optimisation turned off (evidence/sprint-2/D1/hindi-smoke.json).

What this demo does not do yet

  • Test model only (SmolLM2-135M-Instruct q4, English): neither the pilot model (Qwen3-1.7B, whose graph alone is over the 1 GB Pages limit) nor any Hindi model is hosted here. The dockit demo hosts Qwen3-0.6B instead.
  • Not measured on 8 GB: every figure on this page comes from an Apple M5 Pro with 48 GB. Windows is not measured either.
  • The embedding model (multilingual-e5-small) is an English-and-multilingual general-purpose model; this page makes no claim about its quality on Indic text.
  • WebGPU is verified only on an Apple-silicon Mac; Windows and the 8 GB no-GPU laptop are not measured.
  • WebLLM rung: not yet (cut from Sprint 4; revisit after the Sprint 5 measurement). The contract and the staging rules describe it, but its adapter and model are not in this build.
  • Memory figures are shown only where Chrome exposes them: the page's JavaScript heap through performance.memory, which is coarse and Chrome-only. Inside the worker they may all be dashes, and there is no GPU-memory figure.
  • Answers are kept short on WebAssembly (at most 24 tokens) so the timing figures come quickly; that is a limit of this page, not of the engine.
  • No threads: GitHub Pages cannot send the COOP and COEP headers, so WebAssembly runs on one thread and is slow.
  • The worker's network control is the egress guard plus the Content Security Policy the worker inherits; GitHub Pages sends no headers, so this page's policy is a meta tag.
  • The Chrome built-in model is used only if you tick the box below and the browser reports it already available; it is Google's model, not hash-verified by us.
  • No service worker: offline after the first load is best effort, not guaranteed.
  • The model manifest is signed by a key generated for this build only (it lived in memory and was never written to disk), so it proves that the manifest and the files came from the same build, not who published them. This page pins the manifest id (dev.airplanemode.runtime-demo-models) and has no expiry, so it keeps working offline.

Your device

Detected locally by detectCapabilities(); nothing is sent anywhere.

Detecting…

Plan

The rungs of the fallback chain, in order. A rung is skipped before anything is downloaded when the device cannot run it.

    Test model

    Checking this build…

    Embed

    Runs multilingual-e5-small (q8, MIT) through the worker host: two short sentences become two 384-number vectors, and the page shows their length, norm and similarity. The text stays in this tab. The model is a second download (about 119 MiB from this site) and every file is hashed in the worker before it is used.

    Prompt

    The prompt is sent to the model in this tab only and is never stored or logged.

    
            

    Load timing and memory

    The same table the Airplane Mode DevTools panel shows for this page: five phase durations of the load under one load id, and the memory figures Chrome exposes. This page calls exposeDevtoolsHook(), so the real panel can read it too.

    1. Clone the repository and run npm run devtools:build.
    2. In chrome://extensions switch on Developer mode and choose Load unpacked: devtools/dist.
    3. Open DevTools on this page and choose the Airplane Mode tab.

    Details and limitations are in devtools/README.md (the DevTools panel section of the README). The extension is not on the Chrome Web Store.

    Nothing to show before a model is loaded.