Meta releases Muse Spark Open, requiring two Mac Studios to run locally
Meta has released the weights for its 816-billion-parameter model. Apple says its memory footprint means local inference must be split across multiple Mac Studios.
Meta released the weights for Muse Spark Open on October 15, 2026, and says the AI model is optimized to run locally across Mac Studio computers using Apple’s MLX framework. Apple says the model cannot fit on a single Mac Studio, making a two-computer setup central to its desktop use.
Meta describes Muse Spark Open as its most capable US open-weight model to date. The model has 816 billion parameters in total, with 42 billion active parameters, according to specifications reported by The Verge and discussed in Apple’s technical post. Meta describes it as natively multimodal and built for coding and agentic work.
Why Muse Spark Open needs two Mac Studios
Apple puts the model’s footprint at about 493GB in the 4-bit NVFP4 format it evaluated. That figure does not include the memory needed for intermediate computations during inference. Apple says even a Mac Studio with the maximum 512GB of unified memory cannot accommodate the model on its own.
Apple describes splitting the model’s layers between two Mac Studios, with each computer holding half and passing outputs to the other. Two M5 Ultra Mac Studios can provide a terabyte of unified memory across the setup, according to Apple. The requirement concerns a combined configuration: it does not mean a single computer needs 1TB of RAM.
For developers seeking to run the released weights locally, that configuration is a substantial hardware constraint. Engadget reported that the weights were available to download from Hugging Face, but Apple had not announced a specific price or shipping date for the 512GB Mac Studio configuration. Apple had said it would be available in late October.
What Apple evaluated with MLX
Apple’s technical post examines local and distributed inference using MLX and RDMA. It evaluates Muse Spark Open at NVFP4 precision and includes comparisons with Mistral Medium 3.5 and Qwen 3.8 using measures such as time to first token, generation throughput and peak power. The hardware and memory figures describe Apple’s own evaluation; they do not establish an independent comparison of Meta’s broader capability claims.
Read in depth: AI
YouTube Music starts labeling songs made with generative AI
Listeners can tap a label on the player page to see how AI contributed to a track. YouTube says it will use partner disclosures and its own detection tools.
Six parents describe tighter rules for children’s social media and AI use
Guardian interviews show how six families are setting limits on social media and AI. A separate U.S. survey shows parents’ views vary sharply by how teens use chatbots.
Sources and context
- Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run itEngadget
- Meta and Apple team up to make Muse AI models run better on MacsThe Verge
- Local and Distributed Large LLM Inference with MLX and RDMA on Mac StudioApple Machine Learning Research
AI-assisted article checked against the listed sources. NewsJaws did not conduct interviews or attend the reported events.