kaptionai/parakeet-ultra-ane

Parakeet Ultra for the Apple Neural Engine (Kaption)

A Core ML conversion of moondream/parakeet-ultra (revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b), packaged for on-device transcription on the Apple Neural Engine in the Kaption desktop app for macOS 15 and later on Apple silicon.

Parakeet Ultra is a post-trained version of NVIDIA's parakeet-tdt-0.6b-v3 by Moondream. This is an independent adaptation: Moondream and NVIDIA did not make it and do not endorse Kaption.

License and attribution

Licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0), the license of the source model.

• Source model: Moondream Parakeet Ultra, https://huggingface.co/moondream/parakeet-ultra, revision 73175eb7aeb0d82f1e2a6b53b3aabc10a90bcd0b, model.safetensors SHA-256 c9608f36d0ab956c14bfcc525479b0746b3b42a56f6949ec85c14eb7466717dc (CC-BY-4.0).
• Original model: NVIDIA parakeet-tdt-0.6b-v3, https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3 (CC-BY-4.0).
• Changes made by Kaption: converted to Core ML and compressed (see below). The weights are no longer bit-identical to the source.

Files

File	SHA-256	Bytes
kaption-ane-ultra-pal4g16-20.zip	7ed6ad9876a44236b7e02f44cbb59b0ea02ec28c9cc16ff8015d00e93542beab	333,695,546

Inside the zip (391,580,756 bytes installed; every file is listed with its SHA-256 and size in kaption-ane-package.json, whose own SHA-256 is 5fa4eae272f0ad4fb3c637841d10f5000927f5f029e5ab3695995f33dff5ae1e):

• enc.mlpackage/ — the FastConformer encoder (with the encoder projector folded in) as a Core ML ML Program, one static 20-second window, minimum deployment target macOS 15.
• decoder.bin, decoder.json — the TDT prediction network and joint, fp32, decoded greedily on the CPU (Accelerate).
• mel_filters.bin — float32 [128, 257] Slaney mel filterbank matching the source feature extractor.
• tokenizer.json — unchanged from the source revision.
• kaption-ane-package.json — the package manifest.

How it was converted

1. Weights loaded from the source model.safetensors above (all fp16), converted to an fp32 PyTorch FastConformer encoder with the attention rewritten for the Neural Engine layout (numerically equivalent), static 20 s window and masks. The auxiliary VAD head is dropped.
2. Converted with coremltools 9.0 (torch 2.7.0) to an ML Program, then palettized to 4 bits per weight with per-grouped-channel lookup tables (group size 16, "pal4g16") using coremltools.optimize.coreml.
3. Decoder, joint and mel filters exported as raw fp32 arrays for a Swift/Accelerate runtime.

On an Apple M1 (16 GB) with FLEURS (36 clips) the package ran at about 80x real time with WER 3.85% (fp16 Core ML: 3.87%), using about 0.37 GB of system wired memory plus about 0.2 GB in the helper process. Kaption checks every download against the SHA-256 and size above and qualifies the compiled model on the device before using it; anything that fails falls back to Kaption's CPU model.

Use

This package is made for Kaption's kaption-ane helper and its protocol; it is not a drop-in Transformers or NeMo checkpoint. To use the model elsewhere, start from moondream/parakeet-ultra.