The most volatile frontier in MLX
Video generation and music generation are real, working capabilities on Apple Silicon as of 2026-05 — but they are also the most rapidly changing parts of the MLX ecosystem. A model that's the leader today may be superseded next month; a workflow that runs cleanly on mlx 0.31.x may need adjustment after a release or two. This lesson is intentionally short because the ground is moving.
Video, briefly
Video generation models (text-to-video, image-to-video) are large, memory-hungry, and slower per second of output than image generation. Running them locally on a Mac is feasible at the smaller end (a few seconds of low-resolution video on an M-Ultra) but is rarely the right call vs hosted APIs at higher quality. The MLX-native ports are catching up but lag the hosted state of the art by 6-12 months as a rule of thumb.
If you want to play with local video generation, look for mlx-community uploads of CogVideoX-class models, Wan2 / Hunyuan-class derivatives, or whatever's currently leading the open-weights tier. Expect to allocate hours of generation time for a few minutes of output.
Music, briefly
Music generation has a similar shape. Open-weights leaders (MusicGen and successors) have MLX ports of varying maturity. Quality on Apple Silicon is competitive with PyTorch versions of the same models, but the model bench is smaller and changes more often than the LLM/VLM/STT spaces.
What to do with this lesson
Don't pick a video or music model based on what this lesson says — it'll be stale by the time you read it. Use the lesson as a pointer: open mlx-community, sort by recent uploads, look for video / audio tags, and pick whatever's currently active. The pattern (load, generate, save) carries; the specific model identifiers do not.
The honest take
For most creative use cases as of 2026-05, the hosted state of the art for video and music is meaningfully ahead of what you can run locally on a Mac. Local generation in this space is for privacy-sensitive use, for people who want to experiment with the open-weights leaders, or for very specific quantitative work where having full control over the model matters. If you just want a video or a music clip and don't have a strong reason to keep it local, a hosted API will produce better output faster.