Meta Releases Muse Spark Open With 816B Parameters and Local Mac Support

Meta has released Muse Spark Open, a large AI model whose weights developers can download rather than accessing it only through a company-controlled chatbot or API. The October 15 announcement is a notable return to open-weight releases at the high end of Meta’s model range. It also comes with an important qualification: a downloadable model is not necessarily a model that most people can run on an ordinary computer.

Meta says Muse Spark Open has 816 billion parameters in total, with 42 billion active for a given token, and accepts up to one million tokens of context. It is natively multimodal and aimed at coding, multi-step reasoning and agentic work. Those specifications describe scale and design, not a guarantee that it will outperform every rival in a real task. Meta calls it the most capable US open-weight model to date, a claim that should be tested independently.

Releasing the weights gives developers a way to download the trained model and adapt or host it outside Meta’s own service. A company building an internal coding assistant could, in principle, choose its infrastructure and control where its data goes. Researchers can inspect behaviour more directly, while service providers can compete to offer access. As TechBooky’s guide to open-weight AI explains, that is not the same as providing all training data, code and permissions needed to recreate a model from scratch.

The distinction matters here because Meta is releasing Muse Spark Open under its Muse Model License, which it describes as having limited commercial terms. Developers should read those conditions before treating the release as unrestricted open-source software. The model is available through Hugging Face and cloud model catalogues, while hosted access is offered by several inference providers, including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.

Meta is drawing attention to a local setup using two Apple M5 Ultra Mac Studios and the MLX framework. The company says the machines’ combined unified memory can hold the model and demonstrated coding and video-editing workflows running on-device. That is an intriguing example of what local deployment can do, but it is far removed from downloading an app on a typical laptop. Most developers and smaller businesses are more likely to test a hosted version first.

The release is also a strategic shift in a crowded market. Meta recently introduced Muse Wonder models for longer-running agents, while Chinese and US labs continue to make capable open-weight systems available. For developers, choice matters as much as a leaderboard position: a model they can host or customize may be more useful than a marginally stronger one that remains tied to a single API.

Meta says it conducted safety evaluations and added post-training safeguards before releasing the weights. It points developers to a model card for detail, with a fuller evaluation report to follow. Open-weight releases can support independent scrutiny, but they also reduce a provider’s ability to control how copies are deployed. Both sides of that trade-off deserve attention.

The immediate question is whether Muse Spark Open performs reliably and affordably in the coding and agent workflows Meta highlights. Independent testing, deployment costs and the license’s practical limits will tell developers more than a launch-day claim. What is clear today is that Meta has put a very large model into more hands, even if running it privately remains a costly proposition.