Meta unveils Muse Spark Open, an open-weight model of 816 billion parameters

Running an artificial intelligence model of hundreds of billions of parameters on office equipment is an important technical challenge. With Muse Spark Open, presented on October 15, 2026, Meta puts forward an approach that combines a large model with an architecture designed to limit the calculations needed at each stage.

The model has 816 billion parameters in total, but about 42 billion are activated for each token processed. It uses a separate mix-of-expert architecture (MoE): instead of soliciting the entire network for each operation, the system directs the data to some of its specialized experts. This design reduces token calculations, but does not remove the need for memory to load model weights.

Its context window reaches 1,048,576 tokens, or about a million. This ability allows working on very large sets of information in the same session: technical documentation, code files, long documents or history of a task. However, it does not guarantee that every information will always be understood or used correctly.

Meta designs Muse Spark Open for software development and AI agents. With Muse Code, the model can intervene in a multi-step project: review the existing code, prepare a change, apply changes and verify the result. Meta also presents a demonstration in which the DaVinci Resolve pilot system to examine video sequences and select relevant passages. These examples illustrate the intended uses, but are not a guarantee of success on all projects.

One of the most remarkable aspects concerns local execution. Meta explains that it optimized the model with Apple's MLX framework to run on two Mac Studios equipped with M5 Ultra chips. Apple's unified memory architecture allows graphics processors to access a large amount of memory. Two machines can thus distribute the weight storage and the operations necessary for inference.

However, this configuration remains far from a mainstream computer. A model of this size requires a large memory infrastructure, expensive hardware and a suitable software configuration. Local execution can reduce dependence on an external API provider, but it transfers responsibility for equipment, updates, security and maintenance to the organization.

The term "open-weight" also deserves clarification. Meta makes the weights of the model accessible for download, allowing you to host on your own infrastructure and customize it. This does not mean that all training data and all stages of development are published or that its use is without restrictions. Muse Spark Open is distributed under the Muse Model License, the conditions of which must be examined before any commercial use.

For technical teams, the choice between a local model and a cloud service will therefore depend on several criteria: data confidentiality, hardware budget, expected speed, operating costs and quality of results on the tasks concerned. The number of parameters is not sufficient to determine which will perform best in a given context.

Muse Spark Open illustrates an important development: very large open models are no longer only presented as hosting systems in data centres. Specialized local configurations become possible, although their cost and complexity still limit their accessibility. The challenge will be to check whether the control and customization benefits compensate for the resources needed to operate them.