Moonshine

weights

open(safetensors checkpoints on the Hub and .ort files for the library, ungated)

data

described(about 300K hours: public web audio, open datasets named in the paper, and 100K hours prepared internally

code

partial(inference library and LoRA fine-tuning

license

MIT(every Moonshine Streaming model and the English Tiny and Base models, and the code)

superseded-release

legacy non-English non-streaming models(2025 Arabic, Japanese, Korean, Mandarin, Spanish, Ukrainian and Vietnamese Tiny/Base models under the Moonshine AI Community License (free below USD 1M annual revenue), deprecated where a streaming replacement exists

The current Moonshine Streaming models, like the English Tiny and Base models before them, are MIT-licensed and download freely. The training audio is described but not released, and only LoRA fine-tuning is published. Older non-English models remain under a community license that caps free use at USD 1M in annual revenue; they are being replaced by the streaming line, and even read as current they would not change the score.

• https://huggingface.co/api/models?author=moonshine-ai&limit=1000&expand[]=downloads&expand[]=createdAt&expand[]=tags&expand[]=gated recorded 2026-09-27

  Hub listing: moonshine-streaming-tiny, -small and -medium created 2026-01-06 and the non-English streaming models 2026-08-24, tagged "license:mit" and ungated; the legacy non-English models, created 2025, tagged "license:other".

• https://huggingface.co/moonshine-ai/moonshine-streaming-medium/raw/main/README.md recorded 2026-09-27

  Card: "This includes the original Moonshine training sources (about 200K hours of public web data and open datasets) plus an additional 100K hours of internally prepared speech data."

• https://moonshine-voice.readthedocs.io/en/latest/license/ recorded 2026-09-27

  License page: "Where a streaming replacement exists, they are deprecated".

• https://moonshine-voice.readthedocs.io/en/latest/models/domain-customization/ recorded 2026-09-27

  Docs: "The trainer is an opt-in extra."; "moonshine-voice finetune is an alias of python -m moonshine_voice.lora".

• https://raw.githubusercontent.com/moonshine-ai/moonshine/main/LICENSE recorded 2026-09-27

  LICENSE header: "Moonshine models are released under the MIT License by default, in every language and at every size. This includes all streaming speech-to-text models and all English-language models."; "The only speech-to-text models that are NOT MIT are the legacy non-streaming models for languages other than English".

PyPI downloads of the two Moonshine packages, the current moonshine-voice library and the older ONNX package. The Hub checkpoints are downloaded more, but the package is the documented way to run the models.

• https://pypistats.org/api/packages/moonshine-voice/recent recorded 2026-09-27

  "last_month":38846 for moonshine-voice.

• https://pypistats.org/api/packages/useful-moonshine-onnx/recent recorded 2026-09-27

  "last_month":36873 for useful-moonshine-onnx.

Built for live transcription on small devices rather than for the top of the accuracy tables; its own figures on the leaderboard's test sets place it below Whisper's large models, a step below Whisper.

• https://artificialanalysis.ai/speech-to-text recorded 2026-09-27

  The speech-to-text table lists no Moonshine row.

• https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/d2c5b384deccdb82834f41aeaffcc618c00efa2f/english_short_latest.csv recorded 2026-09-27

  Results CSV rows "mistralai/Voxtral-Mini-4B-Realtime-2602,6.4625" (61st) and "AutoArk-AI/Audio8-ASR-0.1B,6.9525" (62nd); no moonshine row.

• https://huggingface.co/moonshine-ai/moonshine-streaming-medium/raw/main/README.md recorded 2026-09-27

  Card table "### Open ASR benchmark results (WER %)" with the row "| **Average** | **12.01** | **7.84** | **6.65** |".