Meta has launched its open-weight model Muse Spark Open, allowing developers to download the model weights for self-deployment and customization, and has partnered with Apple to run the model on two Mac Studios. Meanwhile, Meta is keeping its more capable Muse Wonder series closed-source, seeking a balance between an open developer ecosystem and AI commercialization.

On Thursday, the 15th, U.S. Eastern Time, Meta Platforms, Inc. released what it calls the most capable open-weight model in the United States to date, Muse Spark Open. The model has a total parameter count of 816 billion, supports a 1 million token context window, focuses on coding and agent tasks, and emphasizes long-horizon task execution, complex software engineering, and multi-step reasoning capabilities.

Meta released the model weights for Muse Spark Open on the same day. Developers can access the model through multiple cloud platforms or download the weights to run it on compatible hardware and inference frameworks, and use tools like PyTorch for further customization. The model uses the Muse Model License with limited commercial terms, allowing many developers to use it for free.

Although Muse Spark Open is less capable than Muse Wonder, which Meta launched this Monday, this release fulfills Meta's previous commitment to making its AI technology broadly open to developers. Meta had previously driven the open model ecosystem with its Llama series, but after underperforming in the last round of model competition, Zuckerberg restructured the AI research and development system and shifted toward building more capable closed-source models. This renewed opening of model weights signals that Meta is attempting to find a new balance between expanding its developer ecosystem and retaining cutting-edge technological advantages.

816 Billion Parameters, 1 Million Token Context, Focused on Coding and Agents

In its official blog, Meta stated that Muse Spark Open has a total parameter count of 816 billion, with 42 billion active parameters, a 1 million token context window, and native support for text, images, video, and audio.

Compared to models primarily used for conversation and content generation, Muse Spark Open places more emphasis on long-horizon agent work. Meta says the model has been optimized for complex software engineering, multi-step reasoning, and tasks requiring sustained progress, with the goal of enabling developers to build AI agents that can plan, call tools, and execute tasks continuously.

As in the official blog, Meta AI head and Chief AI Officer Alexandr Wang also called Muse Spark Open the most capable open-weight model in the United States to date on social media. He stated that the model is designed specifically for coding and agent work, such as long-horizon tasks, complex software engineering, and multi-step reasoning, and he reshared the following benchmark comparisons between the model and competing large models published in Meta's blog.

However, "most capable in the United States" remains Meta's own assessment of its product, not an industry ranking that has been comprehensively and independently verified. The competitiveness of an open-weight model depends not only on parameter scale but also on inference performance, operating costs, deployment difficulty, and actual developer experience.

The focus of this release is the model weights, rather than making all training data, the complete training process, and research details public. Developers can deploy and adjust the model themselves within the scope of the license, but commercial use must still comply with the relevant terms of the Muse Model License.

Meta Partners with Apple to Run Large Models on Two Mac Studios

A major highlight of this release is Meta's partnership with Apple, enabling Muse Spark Open to run on two Mac Studios equipped with M5 Ultra chips through the latest MLX framework. Meta's official blog stated that this is the first time leading agent capabilities have been brought to users' desktops through Mac Studios equipped with Apple M5 Ultra chips.

Zuckerberg posted on social media that Muse Spark Open runs excellently on the new Mac Studio, thanked partner Apple and its CEO Ternus, and attached a photo showing the new model running on a Mac computer.

Meta stated that the unified memory architecture of Apple chips is key to enabling local deployment. Each M5 Ultra Mac Studio can provide up to 512GB of unified memory, and two devices working together can accommodate the entire model. Meta says the two machines can work together like a single system with approximately 1TB of unified memory.

Media outlets believe this partnership offers developers a new option: they can run advanced models on their own hardware without relying entirely on renting computing power from cloud servers. For scenarios such as software development, research, and video production, local deployment also helps developers control the model's operating environment more directly.

Meta also demonstrated the model's local agent workflows in its official blog, including using Muse Code to modify an open-source iOS app, fix issues, and complete the build, as well as analyzing video footage frame by frame in DaVinci Resolve, selecting highlight clips, and adjusting aspect ratios.

It should be noted that the model being able to run on two Mac Studios does not mean an ordinary Mac computer can run this complete model in the same way. Meta's demonstration targets high-end hardware configurations, and local deployment still faces limitations such as device costs, memory requirements, and inference efficiency.

For Apple, this partnership also provides a new scenario to showcase the computing power of high-end Macs; for Meta, it helps prove that open-weight models can run on developers' own devices independently of its own cloud infrastructure. The two companies have long competed in areas such as privacy policy, ad tracking, and computing platforms, and this collaboration shows that the AI race is prompting Silicon Valley companies to reassess the boundaries of cooperation.

Weights Freely Open, Multi-Platform Support Lowers Development Barriers

Meta stated that Muse Spark Open is already available through API services on platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Model weights can be obtained from channels such as Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.

In terms of inference frameworks, the model supports SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, and has been optimized for hardware platforms including AMD, Apple, Intel, and Nvidia.

Developers can also use PyTorch and TorchTitan to further fine-tune the model to suit different application scenarios. Meta says the model uses the Muse Model License with limited commercial terms, and many developers can use it for free.

The core of this release approach is to lower the barriers for developers to try, deploy, and customize the model. Developers can first download the weights and experiment locally or on their own infrastructure; when demand grows and larger-scale inference capabilities or additional tools are needed, they can then choose cloud APIs and computing services.

However, open weights do not mean there are no usage restrictions at all, nor does it mean all developers can run the full model at zero cost. Hardware procurement, deployment and operations, and inference computing power may still bring considerable costs, and commercial applications must also comply with the license terms.

Flagship Model Remains Closed-Source as Meta Seeks to Balance Open Ecosystem and Commercialization

Meta revealed on Thursday that Muse Spark Open is less capable than Muse Wonder, which the company launched this Monday. Meta's choice to open the weights of a lower-tier model while keeping its most advanced model closed-source reflects a change in the company's AI strategy.

Alexandr Wang said in an interview with Bloomberg: "For both safety and capability reasons, we generally do not open-source our most powerful models. But we do believe that over time, continuing to open-source excellent models is very important, and we intend to keep doing so."

He also said that open models are crucial for the United States to maintain its advantage in AI competition, stating: "Most AI in the world will run on open models. They should be American."

This statement also places Muse Spark Open in the context of global competition. Chinese companies such as Alibaba and DeepSeek have already gained widespread attention in the open model field, with developers able to adjust models according to their own needs and deploy them across different hardware and applications. Meta hopes to provide a competitive American alternative through open weights and consolidate its influence in the global developer ecosystem.

But the open strategy also carries risks. Once model weights are widely distributed, developers can adjust model behavior to a certain extent on their own, and the model provider's control over the final deployment method decreases accordingly. Meta says Muse Spark Open has undergone additional safety training and was tested according to the company's safety evaluation framework; however, the safety boundaries of open models still depend on how subsequent users deploy and modify the model.

Therefore, Meta has not completely returned to its previous fully open approach, but has adopted a layered strategy: open-weight models are used to expand developer coverage, while closed-source frontier models retain technological and commercial advantages.

Free Models Still Need to Monetize as Meta Targets the Cloud Computing Market

Another significance of Muse Spark Open is that Meta hopes to create demand for its own cloud business and AI services through open models.

Bloomberg reported that Meta is building its own cloud business and plans to sell customers access to AI models and computing infrastructure. Alexandr Wang believes that freely opening Muse Spark Open will not weaken this business, but can instead serve as an entry point to attract developers: users first try the model on their own devices or infrastructure, and as application scale expands, they then pay for more computing power and tools.

This model is clearly different from Meta's traditional advertising business. Open weights themselves do not directly generate subscription revenue, but they can expand the model's usage reach, help Meta build developer relationships, and create commercial opportunities for subsequent cloud inference, enterprise services, and toolchains.

Compared with directly selling paid AI assistants to end users, the commercial returns of open models depend more on ecosystem scale and subsequent service conversion. Whether developers are willing to adopt Muse Spark Open and whether enterprises are willing to pay for related cloud services will determine whether this strategy can generate sustainable revenue.

For Meta, this release is both a signal of its return to open model competition and a test of its AI business model. The company hopes to expand its ecosystem influence through open weights while keeping its most advanced models and high-value-added services within its commercial system. Whether this balance can work will ultimately depend on the model's actual performance, developer adoption rate, and whether Meta can convert its open ecosystem into sustained demand for computing power and services.