Meta launched on Thursday (15) the Muse Spark Open, an open-weight artificial intelligence (AI) model designed for complex programming and reasoning tasks, as well as for the execution of activities by AI agents. The release stands out because it has been optimized to run directly on Apple computers, in collaboration between two companies that have historically competed for space in the technology sector.
According to Meta, the model has 816 billion parameters and a context window of up to one million tokens. The weights are available for download, allowing developers to run and adapt the technology on their own infrastructure, subject to the conditions established by the license.
The initiative also represents a bet on the local execution of advanced models, reducing dependence on AI services hosted exclusively in the cloud.
How does the Muse Spark Open work?
The Muse Spark Open was developed to handle tasks requiring multiple stages of reasoning, complex software projects, and activities carried out by AI agents. Unlike assistants that merely answer questions, these agents can work through sequences of actions to achieve an objective.
The model is multimodal: it can receive information in text, images, audio, and video, although its output is textual.According to Meta’s technical documentation, it uses an architecture called a mixture of experts.
Although it has 816 billion parameters in total, around 42 billion are active for each token processed. This makes it possible to focus processing on specific parts of the neural network without eliminating the memory needed to keep the full model loaded.
Meta says the technology was designed for long-running programming, software engineering, and multi-stage reasoning tasks. The model can also be customized by developers.
Why are two powerful Macs necessary?
- The company worked with Apple to optimize model execution through MLX, a machine learning framework developed for the company’s chips;
- The setup highlighted by the companies uses two Mac Studio computers equipped with M5 Ultra chips, each with up to 512 GB of unified memory. Together, the machines can provide up to 1 TB of memory to keep the model available during execution;
- Apple’s chip architecture allows the integrated GPU to access unified memory. According to Meta, the M5 Ultra’s memory bandwidth reaches 1.2 TB/s, an important feature for feeding the processing of models of this size;
- In practice, this means that running Muse Spark Open locally requires high-performance equipment and is not a viable option for just any Mac. The setup presented is aimed primarily at developers and teams with the resources to build robust infrastructure;
- The company demonstrated the model on tasks that included making changes to an open-weight app for Mastodon, with bug fixes, new features, and project compilation. It also showed the system working with DaVinci Resolve, video editing software, to analyze images and select clips that could be cropped into vertical format.
Read more:
- R$138 million fine prompts Meta to take case to London Supreme Court
- MPF investigates Meta and X over blocking and removal of content in 2026 elections
- Workers’ Party reports Meta to OAS over alleged interference in 2026 elections
Model can be used outside Apple computers
Despite the partnership with Apple, Muse Spark Open is not restricted to Macs. Meta made the weights available for download and announced access through API platforms, including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
The model is also compatible with different inference tools, including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, and llama.cpp. There are configurations for different types of hardware, including accelerators from companies such as AMD, Intel, and Nvidia.
The technical documentation describes versions with different precision formats. The BF16 format requires approximately 1.65 TB of weights, while the compressed NVFP4 version requires about 537 GB.Even the smaller version requires considerable memory infrastructure to run, in addition to the space needed for the weights themselves.
What does it mean to have open weights?
Open-weight models make the trained parameters available so that others can download, run, and, depending on the license terms, adapt the technology. This offers greater flexibility than proprietary systems, whose operation depends exclusively on services controlled by the developing company.
The approach can also make it easier to perform tasks within an organization’s infrastructure, increasing control over data and processes. On the other hand, running large models locally requires hardware, technical expertise, and maintenance resources.
However, making the Muse Spark Open weights available does not mean that the model is entirely open source or that there are no restrictions. Meta uses the Muse Model License, which establishes specific conditions for commercial use.
With the launch, the company expands its investment in models that developers can run and adapt, while the partnership with Apple demonstrates how high-performance personal computers can be used for AI tasks that were previously often associated with large server infrastructures.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and is currently a writer at Olhar Digital.
