Whisper browser model download sizes

Sources: tiny ONNX file list, base ONNX file list, small ONNX file list. Checked 2026-10-30.

What is counted

WebAssembly uses the quantized encoder and merged decoder. The GPU configuration uses a full-precision encoder with a quantized merged decoder. Totals above are the sums of those weight files in bytes. The app displays decimal megabytes, calculated by dividing by a million. Tokenizer, configuration and runtime files are additional downloads. Network transfer can also vary with caching and transport behavior.

A model is pinned to a repository revision so the configuration does not silently change when a remote repository is updated. The metadata file in this project preserves that revision alongside the checked date and source URL. A new model revision needs new testing and fresh size verification.

Storage is not working memory

The size of downloaded weights does not equal the memory needed to run a model. Audio decoding, intermediate arrays, model state and browser overhead also consume memory. Long audio can be much larger after decoding than its compressed file. A successful Tiny run does not establish that Small will fit on the same device.

Start with a short sample and leave the tab open while inference runs. If the browser stops or the GPU fails, try Tiny with WebAssembly. The automatic setting attempts GPU processing when available and can fall back if model initialization fails. It does not promise GPU support on every browser or operating system.

The browser caches model files when permitted. Browser eviction, private browsing or manual cleanup can remove them. Delete all local data in the app to remove saved recordings and the model cache. Export transcripts you want to keep before doing so. Cached weights alone do not install the website as a fully offline application.

Word-timestamp transcription downloads

The original table above describes the segment-timestamp files still used by Translate to English. Transcription now uses the timestamp-enabled exports below, which expose the attention output needed to align recognized words. These are alternative exports of the same Whisper models, not additional models downloaded for one transcription.

Sources: tiny timestamp-enabled file list, base timestamp-enabled file list, small timestamp-enabled file list. Checked 2026-10-30. The app displays the appropriate download size before you start. Word times are model estimates; listen back before treating a boundary as exact.