Meta and Apple Launch Muse Spark Open for Local Desktop Inference

Key points

• Meta released the weights for Muse Spark Open, described as the most capable US open-weight model to date, on October 15, 2026.
• The model is a mixture-of-experts architecture with 816 billion total parameters but only 42 billion active per token, allowing for efficient inference.
• Apple and Meta collaborated to optimize the model for the M5 Ultra chip, utilizing its 1.2 TB/s memory bandwidth and unified memory architecture.
• Running the full model requires a cluster of Mac Studios; a single machine cannot hold the entire model in its native precision.
• The model supports a 1-million token context window, enabling it to process entire codebases and long-horizon agentic tasks locally.

Background

The release of Muse Spark Open follows a trend of large language models being optimized for local hardware to reduce latency and privacy concerns. Previous iterations of Meta’s Muse models focused on cloud-based inference, but this version explicitly targets desktop environments. Apple’s recent hardware updates, specifically the M5 Ultra chip with dedicated GPU Neural Accelerators, provided the necessary infrastructure for this collaboration. The model is available through various API platforms and can be downloaded from Hugging Face, indicating a broad distribution strategy beyond just Apple hardware.

How outlets are covering it

The Financial Times headline highlights the strategic partnership between Meta and Apple, framing the release as a major industry collaboration. Apple’s research team emphasizes the technical performance, noting that the M5 Ultra’s Neural Accelerators provide up to 4.6x faster time-to-first-token compared to the previous M3 Ultra generation. They detail how the model can be distributed across multiple Mac Studios using RDMA over Thunderbolt. Meta focuses on the model's capabilities, describing it as the most capable US open-weight model and highlighting its suitability for agentic coding and multimodal tasks. NVIDIA, in a separate developer blog, positions the same model for its GB300 NVL72 data center hardware, emphasizing high throughput and low latency for enterprise deployments, contrasting with the local desktop focus of Apple and Meta.

Why it matters

This release demonstrates the viability of running frontier-class AI models on consumer-grade desktop hardware, reducing dependence on cloud infrastructure. It signals a shift toward local AI capabilities for developers and enterprises, offering privacy and latency benefits. The collaboration between Meta and Apple also indicates a growing ecosystem for optimized AI hardware, where software and hardware are co-designed for maximum performance. For developers, this means access to powerful models without the costs and security risks associated with cloud APIs.

What to watch

Developers can download the Muse Spark Open weights from Hugging Face and run them using the MLX framework on Mac Studio clusters. Meta has also made the model available through various API providers, including Baseten and Fireworks AI. Apple and Meta are likely to continue optimizing the MLX framework for future hardware generations. NVIDIA is promoting the model for its data center GPUs, suggesting a dual-market strategy for the same model. Users should expect further updates to the model’s safety features and performance benchmarks as more data is collected from real-world usage.