WindstormLabs/windy-lite-ggml

Windy Lite — ggml q5_1 (mobile build)

Windstorm Labs is the origin of record for this artifact. It is created here and promoted downstream; it is never edited anywhere else.

whisper.cpp / whisper.rn build of Windy Lite, quantized to q5_1 so it runs on phones and tablets.

File	ggml-windy-lite-q5_1.bin
Size	59.7 MB
SHA-256	53aaa0467a2ecfe93692a870a41c52394bc3f17955245553fab66e3b05621ccf
Quantization	q5_1
Languages	multilingual
Certified WER	9.06% (grade A)

Lineage

Stage	Value
Upstream base	openai/whisper-base (74M params, MIT)
Windy fine-tune	LoRA, merged — Clinic patient windy-lite
Converted from	WindyWord/listen-windy-lite @ 11ef4da092108720ac3a3df147d58b326e899677
Converter	whisper.cpp @ 2ca53bb45e38, models/convert-h5-to-ggml.py -> whisper-quantize q5_1
Built on	Veron-1-5090, 2026-08-01

Promotion path

WindstormLabs (here, origin) -> WindyWord (product) -> WindyWordMobileApp (mobile).

Downstream copies are push-only. Fix weights here, never downstream.

Honesty note

Distinct weights from upstream (different SHA-256). On short clean English clips the smaller models can be indistinguishable from stock Whisper — a formal bake-off has not been run. Do not claim quality gains without one.

Clinic patient file: sneakyfree/Windy-Clinic -> stt-models/windy-lite.json

Licence

Upstream base is apache-2.0 on Hugging Face, verified against the source repo we converted from. Note OpenAI's Whisper GitHub repo states MIT while its HF model cards state Apache-2.0; we follow the HF distribution point, since that is what we actually pulled. Windy weights are a derivative and are distributed under the same terms, with attribution upstream.