Meta has released Muse Spark Open, an 816-billion-parameter open-weight multimodal model featuring a one-million-token context window. Distributed freely via Hugging Face and optimized for Apple’s M5 Ultra chip through a new MLX framework, running the system locally requires pairing two Mac Studio units equipped with a combined 1 TB of unified memory, hardware representing an investment exceeding 13,000 euros, as reported by punto-informatico.it.
Architecture and Benchmark Performance of Muse Spark Open
The model architecture deploys 816 billion total parameters with 42 billion active parameters per token routing. It handles text, images, video, and audio natively. Meta designed the system specifically for complex software engineering and multi-step agentic reasoning. In Menlo Park benchmark disclosures, the open-weight release outperforms competing architectures including Mistral Large 4, GLM-5.3, and Kimi K3 across coding and reasoning evaluations. Inference execution runs across multiple engines. These include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, providing hardware pathways for AMD, Apple, Intel, and NVIDIA silicon.
Muse Spark Open — weights releasing today — runs great on new Mac Studios thanks to work with @johnternus and Apple. pic.twitter.com/U2wJNrKBOg
Local execution on Apple hardware requires the newly released MLX framework. Mark Zuckerberg demonstrated the setup on X, confirming that the weights release makes local execution viable on new Mac Studios. Developers deploying the model locally face significant hardware costs. Two M5 Ultra Mac Studio units push capital expenditure past the 13,000-euro threshold.
The Meta and Apple Hardware Partnership
The collaboration between Meta and Apple marks a notable shift following years of corporate friction. Meta previously absorbed multi-million-dollar losses after Apple deployed App Tracking Transparency with iOS 14.5, a privacy feature that restricts ad-tracking without direct user consent. Despite past advertising disputes, joint engineering efforts have successfully tailored Muse Spark Open to the M5 Ultra chip.
Current integration remains hardware-specific to the Mac Studio desktop lineup. However, reporting by Axios indicates that Meta models could factor into future iterations of Apple Intelligence.
