Meta releases the open-weight Muse Spark Open

Background

Apple and Meta had worked together on the model, according to Alexandr Wang, who said the collaboration lasted several months.^[1] Muse Spark Open is an open-weight model, which means its trained weights are published for anyone to download and run, subject to the licence.

What happened

On 15 October 2026 Meta released Muse Spark Open. Meta called it the most capable US open-weight model to date.^[1] The weights were published on Hugging Face, a public platform for sharing model files.^[1]

Key details

Architecture and size. Muse Spark Open is a natively multimodal model with 816 billion total parameters and 42 billion active parameters. It has a 1-million-token context.^[1] Active parameters are the ones used for each token. Total parameters include all of the model's parameters, including those not used for a given token. This design, in which only part of the model is active at a time, is called mixture-of-experts.

Target uses. The model targets coding and agentic work, including long-horizon software engineering, meaning multi-step programming tasks that run for a long time.^[1] A coding agent might read a codebase, make changes across many files and test them over many steps.

Hardware. Meta optimised the model for Apple's Mac Studio with the M5 Ultra chip, using Apple's MLX framework. Two machines, each with 512 GB of unified memory, can hold the full model.^[1] Unified memory is memory shared by the processor and graphics parts of an Apple chip. MLX is Apple's open framework for running machine learning models on Apple silicon.

Collaboration. Alexandr Wang said Apple and Meta had worked together on the release for several months.^[1]

Impact

Its optimisation for a Mac Studio setup meant that a small group of high-end desktop machines could hold a model of this size, according to Meta's figures, rather than requiring a data centre. Meta's description of it as the most capable US open-weight model to date is the company's own claim.^[1]