Browser support

The Web SDK needs a modern browser with WebAssembly, ES modules, OPFS, and Web Audio/MediaDevices for the corresponding features. Model availability also depends on device memory and storage; support for a browser does not guarantee that every model fits on that device.

Chrome and Edge

Chrome and Chromium-based Edge are the primary production targets. With cross-origin isolation enabled, they support SharedArrayBuffer, pthread WASM, and the CPU llama.cpp and ONNX/Sherpa artifacts.

Acceleration split

Backend	Artifact(s)	Acceleration today
llama.cpp (LLM/VLM)	`racommons-llamacpp` + `racommons-llamacpp-webgpu`	WebGPU-first in BackendWorker when `shader-f16` is available
ONNX/Sherpa (STT/TTS/VAD)	`racommons-onnx-sherpa` + `-webgpu` twin	ORT WebGPU when probe succeeds, else CPU (+ threads). See ONNX_WEBGPU.md.

RunAnywhere.runtime.active / setAcceleration are LLM-scoped. Speech diagnostics live on RunAnywhere.runtime.speech and the sdk.speechAcceleration event. Do not treat a WebGPU LLM badge as speech GPU.

For the full modality matrix (LLM, VLM, STT, TTS, VAD, embeddings, RAG, voice agent, rerank, segmentation, diarization, diffusion), use RunAnywhere.runtime.modalities. See WASM_AND_WEBGPU.md for why WASM and WebGPU are complementary, not alternatives.

Firefox

Firefox supports the CPU WebAssembly path and pthreads when the application is cross-origin isolated. Treat WebGPU as capability-dependent rather than a baseline requirement. Test the specific Firefox version and deployment headers before claiming accelerated support.

Safari and WebKit

Safari requires cross-origin isolation for threaded WASM. It does not support Cross-Origin-Embedder-Policy: credentialless, so deploy Cross-Origin-Embedder-Policy: require-corp and ensure every cross-origin resource supplies an appropriate CORS or CORP response. The example includes coi-serviceworker.js as a development fallback; production should send the headers from the origin whenever possible.

ONNX and actionable isolation failures

The ONNX/Sherpa artifact uses pthreads. Before registering it, applications must verify crossOriginIsolated and surface an actionable error when it is false, for example: “Speech features require cross-origin isolation. Configure COOP: same-origin and COEP: credentialless (or require-corp on Safari), then reload.” Do not leave initialization pending or present speech as ready.

await ONNX.register({
  acceleration: 'auto', // WebGPU if browser + ORT EP probe OK, else CPU
  threads: 2,
  requireBackendWorker: true,
});

ONNX.accelerationMode / RunAnywhere.runtime.speech.acceleration are 'webgpu' only when the ORT WebGPU EP append probe succeeds — never merely because a -webgpu WASM twin was present.

See DEPLOYMENT.md for server headers, static assets, CSP, and memory/download guidance. Speech vendor/release: ONNX_WEBGPU.md.