Meta Releases Open Weights for an 816-Billion-Parameter AI Model
Meta made the weights of Muse Spark Open, a large AI model for coding and multistep tasks, free to download; Meta says the full model can run on two Apple Mac Studio desktop computers instead of in a data center.
Meta released Muse Spark Open, an artificial intelligence model with 816 billion parameters, and made its weights free to download. Parameters are the internal settings a model learns during training; weights are those learned values, the part of a model that companies usually keep secret. Releasing them lets anyone run, study or adjust the model on their own computers.
Meta says the model can run on two of Apple's M5 Ultra Mac Studio desktop computers, each with 512 gigabytes of memory, using Apple's MLX software, and calls it the first model in its class to run on a desktop. Apple published its own guide and says four linked Mac Studios can produce more than 36 tokens, or word-pieces, a second.
The model is built for coding and for agentic work, meaning multistep tasks an AI carries out on its own. Although it has 816 billion parameters, only 42 billion are used for each word it produces, which keeps it fast. Meta calls it the most capable American open-weight model to date.
The weights can also be downloaded through Amazon, Google, Microsoft and Hugging Face, and the model is offered by several hosting services.
Menlo Park, Calif. — Meta released Muse Spark Open, an artificial intelligence model with 816 billion parameters, as a free download that the company says can run in full on two Apple Mac Studio desktop computers.
Most of the most powerful AI models in use today are closed: a company runs them on its own servers, and people reach them over the internet. An open-weight model is different. Its weights, the numbers a model learns during training that determine how it responds, are published, so anyone can download the model and run it on their own hardware, study how it works or tune it for a particular job. Parameters are those learned values, and their count is a rough measure of a model's size.
The model is what engineers call a mixture of experts: it is divided into many specialized sections, and for each word or word-piece it handles, only the relevant sections switch on. In Muse Spark Open, 42 billion of the 816 billion parameters are active at a time, which keeps it faster and cheaper to run than its size suggests. Meta describes it as natively multimodal, able to work with more than text, and says it can take in about a million tokens, the word-pieces models read, at once. Meta says it built the model for coding and for agentic work, meaning tasks an AI carries out step by step on its own, such as fixing a bug, building an app and testing it.
According to Meta, the model runs on two Apple M5 Ultra Mac Studios, each with 512 gigabytes of memory, using Apple's MLX software, and is the first model in its class to run on a desktop. Apple's chips let the graphics processor use up to 512 gigabytes of memory per machine, so two machines can hold the entire model and, in Meta's words, "behave like one supercomputer with a terabyte of unified memory." Meta's developer documentation also lists four Mac Studios with 256 gigabytes each as an option. Apple's machine learning team published its own guide the same day, linking Macs over Thunderbolt, and said that on four Mac Studios the model generates more than 36 tokens a second. Meta says it tested the setup by having the model make real changes to Ice Cubes, an open-source app for the Mastodon social network, and by having it edit video in DaVinci Resolve, with all the work running on the machines themselves.
Meta calls Muse Spark Open "the most capable US open-weight model to date." The model is released under Meta's own license, which the company says allows many developers to use it for free while limiting some commercial uses. It is also available through cloud and hosting services including Amazon SageMaker, Google Cloud, Microsoft Foundry, Hugging Face, Groq and Together AI. Meta says the model received extra safety training because open weights can be modified once released, and that it met the thresholds in the company's internal framework for releasing models.