avena-parakeet-ultra (GGUF, q8_0)

GGUF conversion of moondream/parakeet-ultra (Moondream's post-training of nvidia/parakeet-tdt-0.6b-v3; same architecture and tokenizer, better WER on every benchmark group in Moondream's card) for mudler/parakeet.cpp at commit e75de9b (v0.5.0).

Weights CC-BY-4.0 (attribution chain: Moondream Parakeet Ultra → NVIDIA Parakeet TDT 0.6B v3).

Details

• 940,663,680 bytes, sha256 1eb16b5880807db4fc510fe00e719c47eb9d52b0e00595047fd0d682342ce11f
• arch tdt (pure TDT — the aux CTC head is not part of the ultra checkpoint)
• 701 tensors, 219 quantized q8_0 per parakeet.cpp's allowlist (encoder FFN/attention projections, subsampling out projection, joint enc/pred); everything else F32
• parakeet.preprocessor.preemph = 0.97 (NeMo/HF featurizer default that the model was trained and evaluated with; see below)
• vad_head.* tensors excluded (~213K params; Photon-internal VAD head that parakeet.cpp does not read)
• Featurizer buffers (fb, window) and metadata lifted from v3's .nemo (identical architecture; deterministic, non-trainable)

Independent cross-validation: the conversion of the same source with the official parakeet.cpp tensor-writing logic is byte-identical for all 701 tensor blobs, and the inverse HF→NeMo mapping was verified against v3 (which ships both formats) with 0 missing / 0 unexpected tensors and max |diff| = 0. The conversion tool lives in the avena repository at tools/convert-ultra/.

Note on preemph

Some third-party conversions of this model set preemph = 0.0. This artifact uses 0.97: the value the NeMo AudioToMelSpectrogramPreprocessor uses for this checkpoint family, the default of transformers' ParakeetFeatureExtractor (the pipeline the model is published/evaluated through), and the value in NVIDIA's official v3 GGUF. An A/B on clean speech slightly favors 0.97; on real/noisy audio the difference is expected to be larger.

Usage

parakeet-cli transcribe --model avena-parakeet-ultra-q8_0.gguf --input audio.wav --timestamps --json