Meta Releases Muse Spark Open: 816B-Parameter Open-Weight Model Engineered for Apple M5 Ultra Silicon

Meta AI Research has officially released the open weights for Muse Spark Open, recognized as the most capable open-weight artificial intelligence model originating in the United States to date. In close architectural alignment with Apple, Muse Spark Open is purposefully optimized for local execution on high-end desktop hardware utilizing Apple's M5 Ultra silicon.

816B MoE Architecture & 1-Million-Token Context Window

Muse Spark Open leverages a natively multimodal Mixture-of-Experts (MoE) architecture comprising 816 billion total parameters with 42 billion parameters actively routed per token. The model is equipped with an unprecedented 1-million-token context window engineered specifically for persistent agentic workloads, whole-repository codebase transformations, and lossless multi-modal reasoning.

For enterprise datacenter inference, the model incorporates auto-tunable Helion GPU kernels within SGLang, unlocking 1.03x to 3.95x kernel-level acceleration on NVIDIA and AMD silicon alongside a 3% end-to-end throughput uplift on AMD Instinct MI350 accelerators.

Desktop Supercomputing: Dual M5 Ultra Mac Studios via MLX

The defining engineering achievement of Muse Spark Open is its viability outside traditional hyperscaler clusters. Utilizing the open-source Apple MLX machine learning framework, the model executes locally on dual Mac Studio units equipped with M5 Ultra processors.

Apple's unified memory architecture allows the M5 Ultra's integrated GPU to address up to 512 GB of shared memory per machine. Interconnecting two Mac Studios yields a unified 1-terabyte memory footprint delivering 1.2 TB/s of memory bandwidth, providing the exact capacity required to hold the complete 816B weight matrix in active memory.

Autonomous Agent Evaluations: Muse Code & DaVinci Resolve

Meta demonstrated end-to-end agentic workflows executed entirely on-device. Paired with 'Muse Code'—Meta's AI coding terminal—the model successfully autonomously triaged, patched bugs, engineered new capabilities, and executed full binary compilations on the open-source Ice Cubes iOS Mastodon repository.

Additionally, the model demonstrated autonomous mastery over DaVinci Resolve video production pipelines, evaluating raw footage frame-by-frame to isolate optimal narrative segments and execute vertical social reframing crops without operator intervention.

“The future of intelligence belongs to everyone, including the sovereignty to run leading-edge models wherever you choose—right on your desktop. Muse Spark Open demonstrates that personal supercomputing is now a reality.”

This milestone accelerates the strategic shift toward open-weight AI sovereignty, enabling researchers and engineers to execute frontier-grade agent reasoning within private workstation perimeters.