On October 15, 2026, Meta officially announced the Muse Spark Open model and released its full weights to the community, declaring it the most capable open-source model developed by an American company to date. What surprised the tech world even more was the direct technical collaboration between Meta and Apple, allowing this colossal model to run locally on desktop computers instead of relying on cloud server infrastructure.
Muse Spark Open is a natively multimodal model featuring a total of 816 billion parameters, with a Mixture of Experts (MoE) architecture that activates only about 42 billion parameters per forward pass. The model supports a context window of up to 1 million tokens and is positioned by Meta to handle extended tasks, complex software engineering challenges, and multi-step logical reasoning sequences.
Optimizing distributed inference with Apple’s MLX framework
Typically, a model exceeding 800 billion parameters requires dedicated server clusters powered by high-end GPUs to load and run inference smoothly. However, Meta teamed up with Apple engineers to bring the model to desktop hardware through MLX, an open-source machine learning framework developed by Apple specifically for Apple Silicon chips.
According to technical documentation simultaneously released by Apple, the recommended setup to run Muse Spark Open locally is a cluster of two Mac Studio machines equipped with M5 Ultra chips. Each machine packs 512 GB of unified memory, creating a combined pool of 1 TB of unified memory. Enabled by MLX’s latest update, the two Mac Studio systems can link directly to distribute computational workloads and share KV cache during inference without hitting bandwidth bottlenecks.
The leaders of both tech giants took to social network X to publicly back the partnership. Meta CEO Mark Zuckerberg expressed his gratitude to Apple’s engineering team, while new Apple CEO John Ternus praised the Mac Studio’s ability to efficiently leverage hardware resources to run a model of this magnitude right on the desks of individual users and research teams.
The trend of bringing frontier AI to local hardware
The debut of Muse Spark Open highlights an important strategic shift among major tech companies: returning control of AI models to enterprises and independent researchers. The ability to run large-scale models on-premises addresses the primary concern of financial institutions, healthcare providers, and government agencies—ensuring absolute data privacy and eliminating the risk of leaking sensitive information via cloud computing pipelines.
Even so, “running on a desktop” in this context still comes with a substantial financial barrier. A dual-Mac Studio M5 Ultra cluster with 1 TB of RAM costs tens of thousands of dollars, placing it well out of reach for the vast majority of mainstream users. While this collaboration demonstrates that Apple Silicon and the MLX framework have matured into a formidable platform for professional AI engineers, bringing 800-billion-parameter models to general consumer devices remains a long road ahead.