On October 15, Meta released Muse Spark Open, opening the weights of its Muse Spark model for download and calling it the most capable open-weights model from a US lab to date.
Key points
- Scale and capability: an 816B-parameter MoE architecture that activates only 42B per token, a context window of 1 million tokens, native support for text, image, audio and video input, and training aimed at coding and long-running agent tasks.
- Benchmark performance: it scores 61.8 on the JobBench agent benchmark, far ahead of comparable US open models; on pure coding benchmarks, however, Chinese open-source models still lead.
- Moving to the desktop: optimized in collaboration with Apple, the model can run locally on two M5 Ultra Mac Studios through the MLX framework; together the two machines act like a supercomputer with 1TB of unified memory, and for the first time a frontier open model truly leaves the data center.
The weights are available on Hugging Face and other channels, and the model can also be accessed through several cloud API platforms. It is released under the Muse Model License, which restricts commercial use.
Source: read the original. This text is a summary, for reference only.
