File	Quant	Size	Notes
`moonshine-tiny.gguf`	F32	104 MB	Full precision
`moonshine-tiny-q8_0.gguf`	Q8_0	33 MB	High quality
`moonshine-tiny-q4_k.gguf`	Q4_K	21 MB	Best size/quality tradeoff

All variants produce correct transcription on test audio.

Model details

Architecture: Conv1d stem + 6L transformer encoder + 6L transformer decoder (288d, 8 heads, partial RoPE, SiLU/GELU)
Parameters: 27M
Languages: English only
WER: 4.55% (LibriSpeech clean), 11.68% (Other)
Performance: 11.2x realtime on CPU (F32)
License: MIT
Source: moonshine.cpp (MIT)

Usage with CrispASR

bash

./build/bin/crispasr -m moonshine-tiny-q4_k.gguf -f audio.wav
./build/bin/crispasr --backend moonshine -m moonshine-tiny-q4_k.gguf -f audio.wav -osrt

Upstream model: UsefulSensors/moonshine-tiny — published by UsefulSensors.
Upstream licence: mit. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.