Key points
- Meta released Muse Spark Open, an open-weight model with 816 billion parameters and a 1 million token context window, available now through several API platforms.
- Meta says it runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework; Engadget, via TechCrunch, reports it needs at least 1 TB of RAM.
- Meta says it tested the model in agentic workflows, including coding fixes for a Mastodon client and reviewing and cropping footage in DaVinci Resolve.
- Meta and Apple worked together on the launch, and Meta CEO Mark Zuckerberg and Apple CEO John Ternus publicly praised each other's work.
Why it matters for marketers: Marketers and content teams get a path to run a large agentic model on their own hardware, which matters for cost, data control and workflow automation.
Meta has released Muse Spark Open, an open-weight model it calls the most capable of its kind in the United States. It is natively multimodal, has 816 billion parameters and a 1 million token context window, and is aimed at complex coding, multi-step reasoning and agentic tasks. Weights are out now, and Meta says the model is available through several API platforms.
The unusual part is the hardware story. Meta says Muse Spark Open can run on a desktop using two Apple M5 Ultra Mac Studios with Apple's MLX framework. The two companies, usually rivals, worked on the launch together for several months, according to Axios as cited by The Verge. Apple also published a technical post about MLX and the model. Local use is not cheap: Engadget, as cited by TechCrunch, reports a requirement of at least 1 terabyte of RAM.
For marketers and content teams, the draw is the open-weight approach. TechCrunch notes such models can be more cost-efficient and offer more transparency and flexibility than closed ones. Running a model locally could also keep sensitive campaign or client material on your own machines, though the hardware bar makes this realistic mainly for agencies and larger teams. Meta says it tested the model in agentic workflows, including coding fixes for a Mastodon client and reviewing and cropping footage in DaVinci Resolve. Those are Meta's claims, based on its own evaluation results.
What to watch: independent benchmarks against other open-weight models, real-world memory and speed figures, and whether smaller configurations appear. Teams curious about local agents can start by testing the model through one of the API platforms before deciding whether the hardware investment makes sense.
Sources
This brief is our own summary. Read the full reporting at the source.
