Meta has released Muse Spark Open, a new language model now available on Hugging Face under the name Muse-Spark-Open-816B-A42B. The naming convention points to a mixture-of-experts architecture with 816 billion total parameters, of which only 42 billion are activated per inference pass. This design, already adopted by several labs to cut compute costs while preserving model capacity, allows performance close to a much larger dense model at a fraction of the runtime expense.
What stands out in this release is its hardware positioning. According to 9to5Mac, Meta is marketing Muse Spark Open as specifically optimized to run on Mac Studio, Apple's workstation equipped with unified memory that can scale up to several hundred gigabytes. This choice makes sense given the MoE architecture itself: since only a fraction of parameters are activated at each step, memory bandwidth becomes the limiting factor rather than raw compute power, which aligns well with the characteristics of Apple Silicon and its high-throughput shared memory.
This direction marks a notable shift in Meta's approach to distributing open models. Previous open-weight releases in the Llama family were largely aimed at cloud infrastructure built around Nvidia GPUs. By explicitly targeting high-end consumer hardware, Meta appears to be extending local inference to powerful desktop machines, without requiring a dedicated GPU cluster. For developers and researchers who own a Mac Studio, this opens the possibility of running a very large model without depending on an API or cloud service.
The release comes with a technical blog post on Meta's developer site, though precise details on benchmarks, pretraining context, or intended use cases remain sparse at this stage. Independent community testing will be needed to confirm whether the claimed Mac Studio performance holds up against conventional GPU-based setups, particularly in terms of generation speed and latency.