Google LiteRT.js

A dated reading of what is claimed, reported, and independently verified in the current evidence.

Current reading

Readiness band and full integer triple

AI-assisted assembly · derived results

Readiness band

Emerging

Primary summary from verified readiness

Evidence strength Growing

Decision

Agent-assisted assessment · independently reviewed · 2026-10-01

Suitable for a controlled browser-inference pilot with validated models and CPU fallback.

Why: Working browser integrations exist, but tests report model-specific GPU correctness failures and substantial compatibility constraints.
Next: Pin the runtime version, compare decoded outputs against a trusted reference, test repeated inference on target devices, and route failing graphs to WASM before expanding deployment.

Constraints

Model-specific WebGPU output errors require end-to-end correctness testing.
Browser, GPU and driver differences require device-specific validation.
WebNN acceleration remains experimental and configuration-dependent.
Large models can exceed the runtime's documented 2GB WebAssembly memory limit.
Unsupported accelerated operations can require partial delegation or full CPU fallback, depending on JSPI availability.
Model conversion, preprocessing, postprocessing and tensor cleanup remain integration responsibilities.
Shader warm-up, model downloads and browser isolation requirements can complicate interactive applications.

Evidence summary

AI-assisted assembly

Counts and dates only. Raw signals, private excerpts, trust records, and internal corpus material are not published here.

Publication record

Revision 4