Jimmi42/Parakeet-V3-MLX

Parakeet‑TDT 0.6B v3 (MLX) — Model Files

This repository hosts the MLX model files (config + weights + tokenizer) for Parakeet‑TDT v3. It is intentionally files‑only (no widget, no runnable code). Use these files with your own codebase or with the parakeetv3_mlx package in your project.

Contents

• parakeet-tdt-v3-mlx/config.json
• parakeet-tdt-v3-mlx/model.safetensors
• parakeet-tdt-v3-mlx/tokenizer.model
• parakeet-tdt-v3-mlx/tokenizer.vocab
• parakeet-tdt-v3-mlx/vocab.txt

How to use

Option A: Download programmatically

from huggingface_hub import snapshot_download
from parakeetv3_mlx.utils import from_pretrained
from pathlib import Path

local_dir = snapshot_download("Jimmi42/parakeetv3-mlx-files", allow_patterns=["parakeet-tdt-v3-mlx/*"])
model_dir = Path(local_dir)/"parakeet-tdt-v3-mlx"
model = from_pretrained(str(model_dir))
res = model.transcribe("audio.wav")
print("".join(t.text for t in res.tokens))

Option B: Download manually

1. In “Files and versions”, download the files under parakeet-tdt-v3-mlx/ into a local folder.
2. Load with:

from parakeetv3_mlx.utils import from_pretrained
model = from_pretrained("/path/to/parakeet-tdt-v3-mlx")

Notes

• Requires the parakeetv3_mlx Python package (your app or local project) and Apple’s MLX.
• Audio: mono 16 kHz WAV recommended (librosa can resample automatically).
• Long audio: enable local attention + chunking in your code for best memory/perf trade‑off.

Benchmarks (Apple Silicon)

• Settings: chunk=120s, overlap=15s, local attention (256,256), bf16
• Device: M4 Pro

Audio (1h)	Wall time	RTF
English	43.6 s	82.6×
German	59.6 s	60.4×

On M4 Max, throughput is typically ~2× higher under the same settings.

About the author

Profile: https://huggingface.co/Jimmi42