ckrivan/parakeet-ultra-coreml

Parakeet Ultra — Core ML (4-bit encoder)

Core ML conversion of moondream/parakeet-ultra, Moondream's post-trained version of NVIDIA's parakeet-tdt-0.6b-v3, for iOS 18+ / macOS 15+.

It uses the same three-package layout, tensor names and shapes as mweinbach1/parakeet-tdt-0.6b-v3-coreml, so anything that runs that build (for example parakeet-coreml-swift) runs this one by changing the repo id.

Package	Inputs	Outputs
encoder.mlpackage (4-bit k-means palettized, convs fp16)	input_features [1,3000,128] f32, attention_mask [1,3000] i32	encoder_hidden [1,375,640], encoder_mask [1,375]
decoder.mlpackage (fp16)	input_ids [1,1] i32, hidden, cell [2,1,640]	decoder_hidden [1,1,640], next_hidden, next_cell
joint.mlpackage (fp16)	encoder_frame [1,640], decoder_state [1,640]	token_logits [1,8193], duration_logits [1,5]

tokenizer.json is Moondream's (identical to v3's). Input is v3's 128-bin log-mel at 16 kHz, 10 ms hop, in 30 s windows. Moondream's voice-activity head is not included.

How it was made

convert.py (in this repo) loads the transformers ParakeetForTDT checkpoint, wraps the encoder with its projector, the LSTM decoder with its projector, and the TDT joint, converts each with coremltools 9, and palettizes the encoder to 4 bits. Run on nvidia/parakeet-tdt-0.6b-v3, the same script produces a transcript identical to the mweinbach1 build, which is how the contract was checked.

Credits and license

CC-BY-4.0, as the originals. Model by NVIDIA (Parakeet TDT 0.6B v3), post-training by Moondream. This repo only converts the format. Accuracy figures are Moondream's and FluidInference's; see their cards.