runntime-whisper

Safetensors weights for Whisper small, repacked for runntime, which runs them on the GPU in the browser through WebGPU.

Upstream: https://huggingface.co/openai/whisper-small
runntime with Whisper: https://github.com/chmjkb/runntime/tree/@chmjkb/whisper-small

Layout

<size>/<precision>/model.safetensors
size precision download
small int8 298 MB
small f16 484 MB

tokenizer.json at the repo root is the upstream one.

  • f16: every tensor cast from the upstream f32 to f16. Greedy transcripts match Hugging Face transformers token for token on our 17 test clips in English, German, French, Spanish, Italian, Polish and Portuguese.
  • int8: the linear-layer weights quantized to int8 in groups of 32 inputs, one f16 scale per group (runntime quantize --bits 8 --f16). The token and position embeddings and the conv weights stay f16. 15 of the 17 test clips match transformers exactly; the other two differ by two commas and one word.

Modifications

These are not the upstream files. Repacked on 2026-11-18:

  • cast to f16, and for int8 quantized as described above
  • laid out as <size>/<precision>/ for the runntime loader

No weight values were retrained or fine-tuned.

License

apache-2.0, inherited from openai/whisper-small. Full text in LICENSE.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for chmjkb/runntime-whisper

Finetuned
(3,865)
this model