Meta launches Muse Spark, an open-weight model with 816 billion parameters for code and agents

Meta expands the Muse family with an open-weight model for developers

Meta released an open-weight version of Muse Spark on Thursday (15th), an artificial intelligence model aimed at programming and AI agents. According to a Reuters report published by ET Tech, the company made the trained parameters available so developers can download and modify the model, something proprietary models do not allow.

The difference is structural: proprietary models keep the weights under the exclusive control of whoever trains them, while open-weight models deliver those parameters ready for use and tuning. This does not mean open training source code or public training data, but it is already enough to run the model outside Meta's infrastructure and adapt it to a specific use case.

The numbers behind Muse Spark

Meta detailed the model's technical specifications:

• 816 billion total parameters, with 42 billion active per inference (a mixture-of-experts architecture, which activates only a fraction of the network at a time);
• a context window of up to 1 million tokens;
• runs locally on two Mac Studio computers equipped with the Apple M5 Ultra chip;
• it is also available through some API providers, without Meta specifying which ones.

The mixture-of-experts design explains why a model with 816 billion parameters can run on desktop hardware: only 42 billion come into play at each inference step, reducing the computational cost compared to a dense model of the same total size.

Not the first piece of the Muse family

In August, Meta had already launched Muse Glimmer, described by the source as a smaller, also open-weight, version. The company did not detail Glimmer's parameters in the report, but the sequence suggests a portfolio strategy: a lighter model for lower-resource scenarios and now a larger model, aimed at heavier code and agent workloads.

This sets the Muse line apart from the Llama family, which until now was Meta's main open bet on language models. With Muse Spark, the company now has two fronts of open-weight models running in parallel, each with a distinct purpose.

Zuckerberg's thesis behind the move

According to the report, CEO Mark Zuckerberg has publicly argued that advanced AI should not remain concentrated among a small number of companies. The source does not include a literal quote from him in this specific piece, only noting this position as the rationale behind Meta's model-opening strategy, which had already been applied with the Llama family before Muse.

The report also weighs the other side: open-weight models give companies more control over their own data, but require significant computational resources and technical knowledge to operate at scale. For a model with 816 billion total parameters, this is not empty rhetoric.

What changes for developers in Brazil

For engineering teams here, the announcement opens a few concrete doors:

1. Data control: running the model locally, even on expensive hardware, avoids sending code and prompts to a third-party API, which matters for companies with stricter data policies.
2. 1-million-token context: in practice, this allows feeding the model an entire codebase or long traces from agents that execute multiple steps, without needing to split the context into chunks.
3. Entry cost still high: two Mac Studio machines with the M5 Ultra chip are not trivial hardware, but they are much closer to a product team's budget than a datacenter GPU cluster, which changes who can even consider running the model in-house.
4. An option outside single-vendor dependence: teams that today depend solely on closed APIs gain an alternative that can be audited, tuned and hosted wherever makes sense.

It is worth remembering that being open-weight does not, by itself, solve the problem of operating cost at scale: the report itself notes that this type of model requires technical expertise precisely so it doesn't turn into an infrastructure project that nobody knows how to maintain.

What remains open

The source does not specify the exact license under which Muse Spark is distributed, nor does it provide benchmarks comparing the model with open or closed competitors on code and agent tasks. There are also no details about which API providers already offer access to the model, nor about usage pricing through those platforms.

These are gaps that should be filled in the coming days, as the developer community begins running its own tests and Meta eventually publishes more detailed documentation about the model.