[BUG] NSRangeException crash in libmlx on macOS Tahoe 26.3.1 during Parakeet model initialization

[okkie2]

Description

MLX crashes with a native NSRangeException on macOS Tahoe while initializing a Parakeet MLX model. This happens before inference begins and before Python can catch any exception.

I reproduced it across multiple MLX / mlx-metal version pairs, so this does not appear to be limited to a single recent MLX release.

Environment

• macOS: 26.3.1
• Build: 25D2128
• Machine: MacBook Pro (Mac16,7)
• Chip: Apple M4 Pro
• Architecture: arm64
• Memory: 48 GB

Python:

• 3.12.11

Package versions tested:

• parakeet-mlx 0.5.1
• mlx / mlx-metal:
  • 0.31.1 / 0.31.1
  • 0.30.6 / 0.30.6
  • 0.28.0 / 0.28.0

Reproduction

Minimal script:

from pathlib import Path
from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
result = model.transcribe(Path(".cache/datasets/nl/audio/nl-0000.mp3"))
print(result.text)

Observed behavior

The process aborts during model initialization with:

*** Terminating app due to uncaught exception 'NSRangeException', reason:
'*** -[__NSArray0 objectAtIndex:]: index 0 beyond bounds for empty array'

The stack trace goes through:

• libmlx.dylib
• MetalAllocator
• MLX Metal device initialization

Expected behavior

MLX should either:

• initialize successfully and allow model loading/inference, or
• fail gracefully with a catchable Python exception

It should not terminate the Python process with an uncaught native exception.

Additional notes

• The crash reproduces on Python 3.12 and Python 3.14.
• The crash happens before inference starts.
• parakeet_mlx uses ffmpeg for audio decoding, so this does not appear to be an input-format problem.
• Because the crash reproduces on multiple MLX versions, this looks like a broader MLX/macOS Tahoe compatibility issue.

Version matrix tested

Attempt 1

• Python: 3.12.11
• mlx: 0.31.1
• mlx-metal: 0.31.1
• parakeet-mlx: 0.5.1
• import works: yes
• model init works: no
• inference starts: no
• result: native NSRangeException crash

Attempt 2

• Python: 3.12.11
• mlx: 0.30.6
• mlx-metal: 0.30.6
• parakeet-mlx: 0.5.1
• import works: yes
• model init works: no
• inference starts: no
• result: same native NSRangeException crash

Attempt 3

• Python: 3.12.11
• mlx: 0.28.0
• mlx-metal: 0.28.0
• parakeet-mlx: 0.5.1
• import works: yes
• model init works: no
• inference starts: no
• result: same native NSRangeException crash

[zcbenz]

Does a simple script like below run?

import mlx.core as mx

mx.eval(mx.arange(10))