On October 15, Meta released Muse Spark Open. Unlike Muse Wonder, which we recently covered, its model weights are downloadable. Weights are the learned numbers needed to run a model yourself.
This is not a small chat app, though. The model is mainly designed for coding and multi-step tasks. Apple demonstrates running it with MLX, its open-source machine-learning library, across two M5 Ultra Mac Studios with 512 GB of memory each.
Why two computers? According to Apple, a memory-saving four-bit version takes around 493 GB for the model alone. Computation needs extra working memory. Even one 512 GB machine is therefore insufficient in Apple's test setup.
MacRumors reports that the 512 GB version is not yet on sale and is expected later in October. The model is also offered by cloud providers, whose service prices apply. Its licence restricts commercial use: open weights do not mean unrestricted use.
The release gives more people a chance to study and adapt the model. For ordinary readers, the lesson is simpler: check a model's memory needs and licence before the word “open” becomes a purchasing decision.