Glimpse-Dictation/Parakeet-Ultra-coreml

Parakeet Ultra: Core ML encoder, GGUF decoder and full GGUF

Model files for Parakeet Ultra in Glimpse. Parakeet Ultra is Moondream's multilingual fine-tune of NVIDIA Parakeet TDT 0.6B V3 (25 European languages, same architecture and size).

File	Download size	Use
parakeet-ultra-Q8_0-decoder.gguf	19.5 MB	Apple Silicon: predictor, joint network, tokenizer and metadata
parakeet-ultra-Q8_0-encoder-mf3.mlmodelc.zip	529 MB	Apple Silicon, macOS 15 or later: int8 Core ML encoder with 15 s, 10.24 s and 5.12 s functions
parakeet-ultra-Q8_0-encoder-macos14.mlmodelc.zip	524 MB	Apple Silicon, macOS 14: int8 Core ML encoder, one 15 s function
parakeet-ultra-Q8_0.gguf	740 MB	Windows, Intel Macs and any CPU/GPU path: full Q8_0 model

Apple Silicon downloads the decoder plus one encoder zip; other machines download the full GGUF.

Accuracy

Measured with Glimpse's engine on an M2 Pro (Neural Engine, int8 encoder), full test sets, against Parakeet TDT V3 in the same layout:

	Parakeet TDT V3	Parakeet Ultra
FLEURS, 25 languages (8499 clips), WER	13.00%	10.77%
LibriSpeech test-clean / test-other, WER	2.19% / 4.03%	2.07% / 3.82%
554 s recording, transcription time	1.76 s	1.43 s

Every FLEURS language is better or the same. On Windows (RTX 4000 Ada, Vulkan, full GGUF) FLEURS WER is 12.53% vs 10.43%.

Provenance

Weights: moondream/parakeet-ultra (CC BY 4.0), based on nvidia/parakeet-tdt-0.6b-v3 (CC BY 4.0). The safetensors were repacked into V3's NeMo layout, converted with transcribe.cpp's convert-parakeet.py, quantized to Q8_0, and the <unk> token marked as a control token (the checkpoint emits it for characters missing from the vocabulary). The decoder is extracted with extract-parakeet-decoder.py; the encoders are converted with per-channel int8 weights.

Artifact	SHA-256
parakeet-ultra-Q8_0.gguf	37611717bcb235e7cc9dbdd24e19e055c41adc9fd4dafee878304af4a16bcb1a
parakeet-ultra-Q8_0-decoder.gguf	5a68f1b297c07436aa7c6bef26712d7ee8bb9db6911741fd6225956cd6d3ef44
parakeet-ultra-Q8_0-encoder-mf3.mlmodelc.zip	e5e36aca6827b2c78ad33f61c30b1e1d7b7a6ab0a8f268c6b6fac6f7b4274e1e
parakeet-ultra-Q8_0-encoder-macos14.mlmodelc.zip	63903e2d1dd039017cf02b82208fca2c72f770dfef10c4d67f56b6b438e90fb1

Runtime requirements

Extract either zip next to the decoder GGUF; both unpack to parakeet-ultra-Q8_0-encoder.mlmodelc, which Glimpse-Speech pairs with the decoder by name. The files use the tdt-0.6b-v3 variant profile and need transcribe.cpp at glimpse-hq main (ec5f294d or later) for dictionary boosting of words the model spells differently. The predictor/joint decoder runs on the CPU.

Glimpse

These files power on-device dictation in Glimpse, a free, open-source dictation app for Mac and Windows. The source is on GitHub.