TL;DR — Key Takeaways
-
Meta released Muse Spark Open, an 816-billion-parameter open-weight AI model designed for complex software engineering, reasoning and long-horizon tasks.
-
Meta partnered with Apple to optimize the model for local execution on two connected M5 Ultra Mac Studio computers using Apple’s MLX framework.
-
The model features a 1-million-token context window and uses a mixture-of-experts architecture that activates 42 billion parameters during execution.
Meta Platforms Inc. on Thursday announced the release of Muse Spark Open, calling it the most capable open-weight artificial intelligence (AI) model built in the United States to date.
The launch also marks an unlikely collaboration with often-bitter rival Apple Inc.: It lets developers run the massive 816-billion-parameter model locally on high-end desktop hardware.
Designed specifically for complex software engineering, multi-step reasoning, and long-horizon tasks, Muse Spark Open features a 1-million-token context window.
The natively multimodal model can process text, images, video, and audio simultaneously. Though the system contains 816 billion total parameters, it uses a mixture-of-experts architecture that routes execution through 42 billion active parameters during use, optimizing speed and resource efficiency.
To enable local execution without relying on cloud data centers, Meta worked alongside Apple engineers to adapt the system for Apple Silicon. By leveraging Apple’s open-source MLX array framework, the model is optimized to run across a paired setup of two Apple M5 Ultra Mac Studio desktop computers.
“The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory,” Meta said in a blog post. Apple’s unified memory architecture allows each M5 Ultra chip to utilize up to 512 gigabytes of RAM at bandwidth speeds reaching 1.2 terabytes per second, providing the memory capacity required to hold and execute the model offline.
The partnership represents a striking shift for the two tech giants, which have historically clashed over user privacy, app store fees, and digital advertising policies. The effort reportedly developed over several months, assisted in part by former Apple executive Daniel Gross.
Meta Chief AI Officer Alexandr Wang said his team was “excited for mommy and daddy to be playing nice again,” adding that Apple and Meta had been coordinating behind the scenes for months. Meta CEO Mark Zuckerberg also publicly thanked Apple CEO John Ternus for facilitating the optimization. Concurrently, Apple published a feature on its Machine Learning Research site highlighting how MLX enables local, private research on desktop hardware.
The joint release underscores a broader industry push toward local AI execution, giving developers and enterprise users greater control over data privacy and reducing latency costs associated with remote server farms.
Following August’s release of Muse Glimmer, a smaller model geared toward standard laptops, Muse Spark Open targets high-demand commercial and academic workflows. On industry benchmarks, the system recorded a 61.8 score on JobBench, outperforming competing models like GLM 5.3.
Muse Spark Open is available to download under a limited commercial license. Raw model weights can be accessed through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. For developers seeking hosted deployments, API access is available through platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Hardware support for the 512GB M5 Ultra Mac Studio configuration is expected to roll out commercially later this month.
