Meta Releases 816 Billion-Parameter AI Model That Runs on Multiple Mac Studios

Meta has released Muse Spark Open, an 816-billion-parameter open-weight AI model that can run locally across multiple Apple Mac Studio computers.

Demonstrated with Apple’s MLX machine-learning framework, the model targets coding, complex reasoning and autonomous AI workflows, although its hardware requirements put the full system beyond the reach of most consumer computers.

Announced on October 15, the model has 42 billion active parameters and supports a context window of up to one million tokens. Its mixture-of-experts architecture activates a subset of the model’s parameters during processing, reducing computation compared with a dense model of equivalent total size. The complete model must still be stored in memory, which creates a substantial hardware requirement.

Apple’s published technical tests document the model running across two and four Mac Studio systems using distributed inference. In the NVFP4 quantized format, its weights occupy approximately 493 GB, while the full model in BF16 precision requires more than 1.6 TB for its parameters alone.

The quantized version still cannot run on a single Mac Studio equipped with 512 GB of unified memory because inference requires additional memory for intermediate computations.

Apple’s four-machine tensor-parallel configuration generated output at 36.73 tokens per second with an 8,192-token context. Under the same test conditions, a two-machine tensor-parallel setup reached 24.99 tokens per second.

The implementation uses Apple’s open-source MLX framework, unified memory and Thunderbolt 5 networking to distribute model computation across the computers. These results demonstrate local inference on a high-end multi-machine system, not support for running the complete model on any Mac without additional hardware.

Meta has also demonstrated Muse Spark Open working with its Muse Code coding terminal. The company used the system to modify Ice Cubes, an open-source Mastodon client for iOS, including bug fixes, feature additions and a full build.

Another demonstration used DaVinci Resolve to review video footage, identify suitable moments and crop clips vertically. The demonstrations are company-reported results rather than independent tests of performance across arbitrary projects.

The model supports text, images, audio and video inputs, with text as its output. Meta’s published evaluation results show competitive performance on selected agentic and reasoning benchmarks, but the model does not lead every category in its comparison with other open-weight systems.

Developers can obtain the downloadable weights from the Muse Spark Open repository on Hugging Face. Meta also lists Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models as distribution channels.

The release supports deployment through several inference frameworks, including MLX, SGLang, vLLM and Hugging Face Transformers, with hardware requirements varying by configuration.

Muse Spark Open is released under Meta’s Muse Model License rather than an unrestricted permissive license.

The terms allow use, modification, fine-tuning and redistribution subject to stated conditions, including attribution requirements and a separate-agreement provision for certain model-as-a-service providers whose revenue exceeds $20 million during a consecutive 12-month period.