Music Apple Silicon Benchmarks
How long Stable Audio 3 takes to write a music bed on Apple Silicon, through the same generate_music call an AI assistant makes.
Music generation time per bed
Wall-clock seconds, timed through the same tool an AI assistant calls.
- One music bed
- One generate_music call for an instrumental bed about a minute long, from a written brief. The figures below count each bed as 64 seconds.
- With the model already loaded, a bed takes about five to eleven seconds. The slowest time in the range includes loading the model from disk.
A minute of music takes about eight seconds on a Mac that spends over seven minutes on one five-second video clip. Music runs at roughly eight times realtime and video at about a ninetieth of realtime. Music is the only generator you can call mid-task without planning around the wait.
Yes, only the first. The model loads from disk once and stays in memory, so every later call skips the load. The video model is the opposite: it unloads after every clip and loads again for the next.
Makers
The companies behind these models.
Stable Audio 3 models and download sizes
Every model GroundSlate offers for download, including ones not measured yet.
How music generation works on a Mac
Stable Audio 3 comes as three checkpoints in one repo. GroundSlate installs a small model tuned for music and a small one tuned for sound effects, which share two of their three files. This page times the music tool, generate_music, making a bed about a minute long.
Music is by far the fastest generator GroundSlate runs. A sixty-second bed comes back in seconds, while the same Mac spends minutes on one five-second video clip.
The weights download from Stability's optimized-runtime repo, which holds the MLX conversions, not from the main model repo. Same publisher, different repo.
How these numbers were measured
- Machine
- MacBook Pro, Apple M3 Max, 64 GB unified memory
- Engine
- mlx
- Default model
- stabilityai/stable-audio-3-optimized
- Models locked
- Yes
- Path
- Timed through the generate_music tool, so the figure includes the queue and saving the wav to the library.
Questions
- How does making music compare to making video on the same Mac?
- A minute of music takes about eight seconds on a Mac that spends over seven minutes on one five-second video clip. Music runs at roughly eight times realtime and video at about a ninetieth of realtime. Music is the only generator you can call mid-task without planning around the wait.
- Does the first music request take longer than the ones after it?
- Yes, only the first. The model loads from disk once and stays in memory, so every later call skips the load. The video model is the opposite: it unloads after every clip and loads again for the next.
- Which machines are in this dataset?
- MacBook Pro, Apple M3 Max, 64 GB unified memory (Reference Mac).