Meta released the weights for Muse Spark Open on Thursday and calls it the most capable US open-weight model to date. It’s natively multimodal, 816 billion parameters with 42 billion active, and a one-million-token context window. It’s built for coding and agentic work: long-horizon tasks, complex software engineering, multi-step reasoning.
The Apple part
Muse Spark Open is, by Meta’s account, the first model in its class that runs on a desktop: two Mac Studios with M5 Ultra and the current MLX framework. The arithmetic behind it is Apple’s unified memory, which lets the integrated GPU address up to 512 GB per Mac Studio, so two machines hold the whole model, with 1.2 TB/s of bandwidth to read the small slice each token needs.
CNBC reports that customers can pair two to four Mac Studios, and that the collaboration took months. The notable thing isn’t the engineering, it’s the pairing. Apple and Meta spent years avoiding each other, especially after Apple’s 2021 privacy changes, which Meta estimated in early 2022 would cost it $10 billion in revenue that year. Neither side would tell CNBC whether this leads anywhere further.
Zuckerberg posted a picture on X of the Meta agent running on a Mac, writing that the model runs great on the new Mac Studios thanks to the work with Apple. Ternus replied with congratulations to the team.
Where it lives and what it costs
Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. API access comes via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and Nvidia silicon.
The license is the Muse Model License, with limited commercial terms that Meta says keep it free for many developers. The open-weight version got additional safety post-training, evaluated against Meta’s Superintelligence Scaling Framework. The full evaluation report is still pending.
Open weights are no longer a side show
Meta’s demos aim squarely at the ground Claude Code and its rivals work on: real iOS changes to an open-source Mastodon client, bug fixes, new features, a full build, plus driving DaVinci Resolve frame by frame. All on-device.
For anyone running agents in the cloud today, that shifts the math. Not on capability, where independent comparisons don’t exist yet, but on where the code sits while the work happens. Two Mac Studios are expensive, and they are a one-off purchase rather than a token meter.