Reesthebeast/parakeet-tdt-0.6b-v2-bf16

Reesthebeast/parakeet-tdt-0.6b-v2-bf16

mlx-community/parakeet-tdt-0.6b-v2 with the weights stored as bfloat16 instead of float32. Nothing else differs — same architecture, same tokenizer, same values.

Why

parakeet_mlx.from_pretrained loads with dtype=mx.bfloat16 by default, so the published float32 weights are narrowed to bfloat16 the moment they are read. Storing them that way makes the download 2.47 GB → 1.24 GB with no effect on output: every tensor here is bit-identical to what the float32 repo becomes once loaded.

Verified by transcribing the same audio through both and comparing the text character for character.

Use

from parakeet_mlx import from_pretrained
model = from_pretrained("Reesthebeast/parakeet-tdt-0.6b-v2-bf16")

Built for LocalFlow, a fully-local push-to-talk dictation app for macOS, where this halves a new user's first-run download.