Muse Spark Open

What it is

Muse Spark Open is an open-weight large language model that Meta announced on 15 October 2026. Meta called it the most capable US open-weight model to date.^[1] “Open-weight” means the trained weights are published, so anyone can download the model and run it on their own hardware, rather than using it only through a company's app or interface. Muse Spark, the first model in the Muse family, was different. It launched as a closed model available only through Meta AI and a private API preview.^[3][1]

Origins

The Muse family began with Muse Spark on 8 April 2026.^[4] Muse Spark Open was developed with Apple's involvement. Alexandr Wang said Apple and Meta had worked together on the model for several months, and he credited former Apple executive Daniel Gross with helping to make it happen.^[1]

Key features

The model uses a mixture-of-experts design. In this kind of model the network is made of many specialist sub-networks, and each input activates only some of them, so the work done per step is far smaller than the total size suggests. Muse Spark Open has 816 billion total parameters, of which 42 billion are active.^[1][2]

It is natively multimodal, meaning it was built from the start to handle more than one type of input. It has a 1-million-token context window, which is the amount of text it can take into account at one time.^[1]

Meta focused the model on coding and agentic work. Agentic work means tasks where the model plans and carries out several steps on its own. Meta named long-horizon tasks, complex software engineering and multi-step reasoning as the areas of focus.^[1]

Meta optimised the model for Apple's Mac Studio with the M5 Ultra chip, using Apple's MLX framework. Each machine has up to 512 GB of unified memory, so two of them together can hold the full model.^[1] Meta demonstrated on-device agentic tasks with its Muse Code coding agent, including iOS app bug fixes and video editing in DaVinci Resolve.^[1] The weights were released through Hugging Face, and Apple published a companion article on distributed MLX inference.^[1]

Where it fits

Muse Spark Open is the open-weight member of the Muse family. Its release shows that the family now includes downloadable weights alongside the closed model that first launched it.^[3][1] For readers, it is a concrete example of the open-weight approach: people with suitable hardware can inspect the model, run it locally and build on it, which a closed model does not allow.