Meta Releases Muse Spark Open AI Optimized for Mac Studio

Meta has released Muse Spark Open, a downloadable AI model optimized to run locally across multiple M5 Ultra Mac Studios using Apple's MLX framework. Apple has also published benchmarks showing how the model performs on two- and four-Mac setups.

The multimodal model is designed for coding, AI agents, and multi-step reasoning, with a context window of up to 1 million tokens. It uses a mixture-of-experts architecture with 816 billion parameters in total, activating 42 billion at a time rather than using the entire network for every step.

Running it on a desktop still takes substantial hardware. Meta's optimized setup uses two M5 Ultra Mac Studios, each with 512GB of unified memory. Even a compressed, 4-bit version of the model occupies roughly 493GB before allowing for the extra memory needed to run it, so it cannot fit on a single 512GB Mac Studio.

Apple's M5 Ultra Mac Studio provides 1.2TB/s of memory bandwidth per machine. These demonstrations use the highest-memory configuration, which Apple announced for late October, rather than the base model.

The latest MLX release takes advantage of the M5 GPU's Neural Accelerators and splits the workload across Thunderbolt 5-connected Macs. In its technical breakdown, Apple describes using RDMA, or remote direct memory access, to move data between the machines with low latency. Each pair of Macs in the benchmark setup was connected with two Thunderbolt 5 cables.

Apple measured the following generation speeds for the 4-bit model, using an 8,192-token context and generating 128 tokens per test:

Pipeline parallelism divides the model into sections that run on different Macs, while tensor parallelism shares the work within individual layers. The latter delivered higher generation speeds in these tests, reaching about 25 tokens per second on two Macs and 37 on four. These results measure text generation at the stated context length, not performance across the model's full 1-million-token window.

Meta also demonstrated the model doing more than answering prompts. Using Muse Code, it made changes to the open-source Ice Cubes iOS app, including bug fixes, new features, and a full build. Another demonstration had the agent operate DaVinci Resolve to review footage frame by frame, select moments, and crop clips vertically. Both demonstrations ran the model locally.

The release builds on Apple's earlier work to run AI agents locally with MLX, but Muse Spark Open is not limited to Macs. Meta lists support for other inference engines, including llama.cpp, SGLang, and vLLM, alongside cloud and API providers.

The weights are available through Hugging Face and other distribution partners. They are offered under the Muse Model License, with free use for many developers and limits on commercial use, rather than an unrestricted open-source license.