Meta has released Muse Spark Open, an 816-billion-parameter open-weight AI model that it demonstrated running across connected Apple Mac Studio desktop computers rather than a conventional data-center GPU cluster. For businesses evaluating advanced AI, the announcement raises an important question: Must powerful AI agents remain dependent on cloud infrastructure, or can organizations operate them on computers they control?

Designed for complex tasks involving coding, planning, and software tools, Muse Spark Open gives developers and organizations the ability to download, customize, and deploy a powerful AI model locally. Apple's distributed-inference technology allows the model's workload to be shared across specially configured desktop computers, combining their computing resources. For businesses, these capabilities introduce new choices about AI infrastructure, data handling, and developing specialized applications.

However, bringing advanced AI capabilities to the desktop does not mean making them accessible to every computer or business. Running Muse Spark Open locally requires substantial computing resources, technical expertise, and compliance with licensing conditions. While the model expands deployment options, its practical value depends on performance, infrastructure costs, and the requirements of individual organizations.

Muse Spark Open demonstrates that powerful models designed for AI-agent workflows can run beyond traditional data centers, but widespread local deployment remains a more complicated proposition. The significance of Meta's release lies in what the model can accomplish, how Apple helped make desktop deployment possible, and whether those capabilities justify the investment required to use them.

Meta's Muse Spark Open is an 816-billion-parameter open-weight AI model designed for complex reasoning, coding, and AI-agent workflows. Developers can download, customize, and deploy the model on supported computing infrastructure, expanding their options beyond proprietary cloud-based AI services.

With its October 15 release, Meta made Muse Spark Open's trained weights available for download—the numerical parameters that determine how an AI model processes information and generates responses. With proprietary models such as Anthropic's Claude, developers generally access the technology through hosted services rather than downloading its underlying weights. Muse Spark Open instead allows developers to adapt the model for their own applications and operate it on supported infrastructure, subject to Meta's licensing terms.

That flexibility distinguishes Muse Spark Open from Meta's more powerful Muse Wonder and Muse Wonder Full models. The Wonder models are designed for advanced personal and professional AI agents, with features such as Active Memory and Active Focus that help them retain relevant information and continue working toward complex goals. Meta provides access to those models through its services, including the Meta Model API and Muse Code. Muse Spark Open takes a different approach by making a capable model from the Spark family available for developers to download and customize themselves.

Making the model independently deployable also creates additional safety considerations. Without Meta's hosted safeguards surrounding it, developers must account for potential misuse within their own applications. Meta says Muse Spark Open received additional safety post-training for standalone use and has evaluated risks involving harmful responses, unauthorized actions through software tools, and improper handling of sensitive information. The company nevertheless recommends that developers implement safeguards appropriate to their intended applications.

Meta describes Muse Spark Open as the most capable U.S.-developed open-weight AI model released to date, although that remains the company's characterization rather than an independently established universal ranking. Its combination of capabilities and accessibility also reflects Meta's broader goal of developing what it calls "personal superintelligence"—advanced AI that helps people accomplish complex tasks in their professional and personal lives. By making the model's weights available, Meta is expanding opportunities for outside developers to create specialized applications rather than limiting those possibilities to its own products and services.

Meta and Apple worked together over several months to make Muse Spark Open run across connected Mac Studio computers, bringing two longtime technology rivals together on an AI project. Their collaboration demonstrates another way to deploy a powerful model locally, although the Apple configuration is not the only way to operate the model.

But greater control over AI infrastructure is only part of the equation. For developers and businesses, the value of that flexibility ultimately depends on what the model can accomplish once it is deployed—and whether its capabilities are strong enough to support useful, reliable applications.

Muse Spark Open is designed to do more than respond to questions or generate content. Meta trained the model for long-running AI-agent workflows, where completing a task may require several decisions and interactions with other software. Rather than simply producing instructions for a user to follow, the model can plan a sequence of actions, use connected tools, examine the results, and attempt to correct errors when something goes wrong. These capabilities are particularly relevant to work that involves multiple steps, such as developing software or processing complex information.

Although Muse Spark Open contains 816 billion parameters, it uses a mixture-of-experts architecture that divides parts of the model into smaller neural networks called experts. For each token it processes, the model selects eight of its 256 experts, activating approximately 42 billion parameters rather than the entire model. This reduces the computing power needed to process information, although the full model still requires substantial memory to operate. It also supports a context window of approximately one million tokens, allowing developers to provide extensive materials, such as large codebases or collections of documents, within a single interaction. That capacity makes longer workflows possible, although it does not guarantee that the model will accurately interpret or remember every detail.

Muse Spark Open can also process text, images, video, and audio, giving it access to information beyond written instructions. For example, an AI agent could examine screenshots to understand a software interface or analyze video footage before deciding what action to take. Although the model can only generate text rather than images or video, it can use its understanding of those inputs to guide actions in connected software.

Meta illustrated these capabilities through Muse Code, its AI coding environment, using Muse Spark Open to work on Ice Cubes, an open-source Mastodon application for iOS. According to the company, the model implemented bug fixes and new features and completed a build of the application. The example shows how an AI agent can move beyond suggesting code changes to carrying out steps in a software development process, including working with the tools needed to implement those changes.

The company also showed Muse Spark Open operating DaVinci Resolve, a professional video-editing application. The model reviewed footage, identified notable moments, and used the editing software to create vertically cropped clips. This requires more than following written instructions: the model must interpret visual content, determine which moments are relevant, and use software tools to produce the requested edits. Together, the coding and video examples illustrate how Muse Spark Open can apply its capabilities across different kinds of work, using information it interprets to carry out actions within specialized applications.

Meta's published benchmark results position Muse Spark Open as a competitive open-weight model across AI-agent tasks, coding, reasoning, and visual understanding, although its performance varies by application. In the company's testing, Muse Spark Open scored 89.2 on MCP-Atlas, a benchmark measuring how well AI agents select and use external software tools to complete multistep tasks. That result was higher than those of the competing models in Meta's comparison, suggesting particular strength in workflows that require AI to interact with multiple software systems.

The picture is less clear-cut for coding. Muse Spark Open scored 76.4 on Terminal-Bench 2.1, which evaluates whether AI agents can complete practical computing tasks through a command-line interface. That placed it behind several competing open-weight models in Meta's comparison, including Qwen 3.8 and GLM 5.3. Meta also reported results on difficult reasoning questions and visual-understanding tests, including scores of 26.1 on HLE-Diamond without tools and 73.6 on SimpleVQA with tools. These results show a model with capabilities across several areas, but not one that leads every performance category.

Independent testing provides another perspective. Plotly, a company specializing in data visualization and analytics software, evaluated Muse Spark Open using 43 questions based on a simulated wind-energy operation. The dataset contained approximately 20 million records across 44 tables, including equipment performance, maintenance histories, and operational problems. The questions tested more than the ability to retrieve numbers: some required the model to recognize ambiguous requests, identify faulty information, or find patterns hidden within the data.

Muse Spark Open answered 38 of the 43 questions correctly, achieving an 88% score. That tied GLM 5.3 Flash and Qwen 3.8 27B for the highest result among open-weight models on Plotly's leaderboard at the time of testing. It was also the highest-scoring U.S.-developed open-weight model in that evaluation. Meta's proprietary models performed better, however, with Muse Spark 1.3 scoring 93% and Muse Wonder reaching 95%. These results show that Muse Spark Open was competitive with the leading downloadable models, although it did not match Meta's hosted alternatives in this particular test.

The individual results reveal both the model's strengths and its limitations. Muse Spark Open correctly answered all 20 standard analytics questions and handled all 10 questions involving ambiguous or impossible requests. However, it identified only two of five deliberately hidden data patterns. In another case, the model incorrectly flagged a valid measurement as impossible and recommended changing it. It also drew an incomplete conclusion about whether windy conditions increased equipment alarms because it did not investigate other possible explanations, such as seasonal differences. For businesses relying on AI to analyze operational data, errors like these could mean overlooking an important pattern or incorrectly identifying valid information as faulty, potentially affecting the decisions that follow.

Meta acknowledges similar limitations in Muse Spark Open's model card, warning that the model can make mistakes during complex, multistep reasoning, particularly when encountering unfamiliar situations. This matters for AI-agent applications because an error early in a workflow can affect subsequent decisions and actions. Strong benchmark scores alone do not establish that the model can reliably complete every extended task without oversight.

Safety performance is another consideration, particularly for a model intended to operate through external tools. Meta reports that Muse Spark Open met its release criteria across evaluations involving harmful content, unauthorized agent actions, privacy, and more severe risks. In its cybersecurity testing, the model refused 94.8% of jailbreak attempts across a collection of cyber-misuse benchmarks. Those results reflect Meta's testing of the released model, however, and do not establish that customized versions or applications built around it will behave identically. Meta recommends additional safeguards, including human confirmation before AI agents perform irreversible actions.

Taken together, the results support describing Muse Spark Open as a capable competitor among leading open-weight models, particularly for tool-based workflows and data analysis. Its performance is promising, but the independent testing also shows why completing a task and completing it accurately are not always the same thing. For organizations considering the model, understanding where it succeeds—and where it can make consequential mistakes—is as important as its position on a benchmark leaderboard.

The collaboration between Meta and Apple is notable not only for its technical achievement but also for the relationship between the two companies. Longtime competitors that have publicly clashed over privacy and technology platforms, Meta and Apple worked together for several months to make Muse Spark Open run on Apple's Mac Studio computers. According to Axios, Meta AI chief Alexandr Wang confirmed the collaboration, which gave both companies an opportunity to demonstrate how powerful AI models could operate outside traditional data-center infrastructure.

The demonstration uses two Mac Studio computers, each equipped with Apple's M5 Ultra processor and 512 GB of unified memory. Unlike conventional computer systems that maintain separate memory for the processor and graphics hardware, Apple's unified memory architecture allows both to access the same memory within each Mac Studio. By combining the resources of two computers, the configuration provides approximately 1 TB of memory. Meta also supports a configuration using four Mac Studios with 256 GB of memory each, providing the same combined memory capacity across more computers.

Making this possible required more than connecting the computers. Apple's MLX software distributes the model's workload across the machines, while Thunderbolt 5 connections allow them to exchange information at high speed. Each computer stores part of the model's compressed weights and performs its share of the calculations. Rather than repeatedly transferring the model itself, the computers exchange smaller amounts of intermediate information as they process requests. This allows them to operate together as a single AI inference system, using a model that would otherwise exceed the memory available on one Mac Studio.

Apple's published performance tests provide an indication of how well the arrangement works. Using a configuration with two M5 Ultra Mac Studios, Muse Spark Open generated approximately 25 tokens per second, compared with 36.7 tokens per second when running across four machines. Tokens are the units of text an AI model processes and generates, so the results offer a measure of its response speed. However, these figures come from a specific test using an 8,192-token context and should not be interpreted as guaranteed performance across different applications or workloads. That's because the measurements reflect how quickly the model generated text under those conditions, not the additional time it may take to process larger amounts of information or interact with external software tools.

There is also an important distinction between demonstrating the technology and making it immediately available to customers. Although Apple's new M5 Ultra Mac Studio began shipping on September 22, the company scheduled its 512 GB memory configuration for late October. That meant the featured two-computer setup was not yet generally available when Meta announced Muse Spark Open on October 15. Organizations interested in replicating the demonstration would need the appropriate hardware and technical configuration.

The achievement does not mean an ordinary desktop computer can now run any advanced AI model. Instead, Meta and Apple have demonstrated that a model of this scale can operate across connected professional desktop computers without relying on a conventional server cluster for inference. The collaboration establishes a practical alternative for certain advanced AI workloads, although whether that approach makes financial and operational sense for organizations is a separate question.

Developers and organizations can download Muse Spark Open directly from Meta's official model repository on Hugging Face, where users must sign in and complete the required access process. Meta also provides developer documentation explaining how to install, configure, and deploy the model. For organizations using established cloud platforms, Muse Spark Open is also listed in Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. However, obtaining access to the model is different from having the resources to operate it. Organizations can deploy Muse Spark Open on computing infrastructure they manage themselves or use providers that handle the computing resources on their behalf. Each approach involves different technical requirements, operating costs, and levels of infrastructure responsibility.

Organizations interested in running Muse Spark Open themselves are not limited to Apple's hardware. Meta launched support for inference software compatible with NVIDIA, AMD, Intel, and Apple processors, allowing developers to choose among different computing environments. Additional compressed versions expand those possibilities. For example, Meta's documentation describes an Unsloth version compressed to approximately 2 bits per parameter that can fit into about 300 GB of combined computer memory on a single workstation. While this reduces the hardware capacity required, it does not guarantee the same processing speed or output quality as other configurations.

Hardware costs remain a significant consideration. Apple's Mac Studio with the M5 Ultra processor starts at $5,499 in the United States, but that is the base price, not the cost of a system equipped with 512 GB of unified memory. Replicating Meta's featured two-computer configuration would therefore require a substantially larger investment than purchasing two base-model Mac Studios. Organizations operating their own systems must also account for electricity, maintenance, software integration, and the technical expertise needed to keep the model running.

For organizations that do not want to purchase or manage specialized hardware, hosted inference provides another option. Companies including Baseten, Fireworks AI, OpenRouter, and Together AI announced access to Muse Spark Open through their services. Depending on the provider, organizations can use shared computing resources, pay according to model usage, or arrange dedicated infrastructure. This allows developers to test and integrate Muse Spark Open without purchasing the equipment needed to run it independently, although ongoing usage charges and provider-specific terms still apply.

The model's licensing terms introduce another distinction between accessibility and cost. Meta's Muse Model License permits developers and organizations to download, modify, distribute, and use Muse Spark Open commercially without an upfront licensing fee, subject to its conditions and usage policy. Users must retain required licensing information and prominently display the acknowledgment "Built with Muse Spark Open" in products using the model or in accompanying documentation, websites, or settings.

One provision specifically affects larger companies offering AI models as a service. If an organization and its affiliates generate more than $20 million in combined gross revenue during any consecutive 12-month period, and the organization provides third parties with access to AI models through an API or similar service, it must notify Meta and enter into a separate written agreement before offering Muse Spark Open as part of that business. The restriction does not apply simply because a company exceeds the revenue threshold. Meta explicitly excludes internal use and ordinary commercial products that incorporate the model as a feature rather than offering AI models as a service.

For businesses evaluating Muse Spark Open, the absence of an upfront model licensing fee should not be confused with the absence of operating costs. Purchasing and maintaining hardware may make sense for organizations with sustained workloads and the technical resources to support them, while hosted services may be more practical for those testing applications or managing variable demand. Neither approach is automatically cheaper. The financial decision depends on how frequently the model will be used, the infrastructure required, and the ongoing costs of supporting it. Those considerations also determine which organizations can realistically take advantage of the greater flexibility Muse Spark Open provides.

Meta's release of Muse Spark Open gives organizations another decision to make when adopting advanced AI: not just which model to use, but where it should operate and how much control they want over the technology. For businesses developing AI applications, that choice can affect how they handle sensitive information, manage security, customize software, and plan long-term technology investments.

One potential advantage of running Muse Spark Open locally is greater control over company information. When an organization operates the model on its own infrastructure, prompts and proprietary data can be processed within that environment rather than sent to an external AI provider. However, local deployment does not automatically guarantee privacy or security. Applications connected to outside databases, cloud services, or other software tools may still transmit information externally. Businesses must evaluate the entire system, not simply where the AI model runs.

The ability to customize Muse Spark Open creates additional opportunities. Organizations can adapt the model for specialized applications rather than relying exclusively on the capabilities and restrictions of a hosted proprietary system. However, successful deployment requires more than selecting a capable model. Businesses must test its performance against their actual workflows, establish appropriate permissions, and determine when human review is necessary. Meta's demonstrations and independent evaluations show promising capabilities, but they do not establish that Muse Spark Open can reliably complete every business task without oversight.

Local deployment also raises questions about infrastructure and environmental impact. Running AI on computers outside traditional data centers could change where computing resources and electricity are consumed, but it does not necessarily reduce total energy use or water consumption. Those outcomes depend on the efficiency of the hardware, how frequently it operates, and the infrastructure supporting it. Meta and Apple's demonstration does not establish that desktop deployment offers an environmental advantage over centralized computing.

In the near term, Muse Spark Open is likely to be most practical for developers, AI infrastructure teams, and organizations with the resources to manage advanced computing systems. Businesses without those capabilities can still evaluate the model through hosted services rather than committing immediately to their own infrastructure. Some organizations may find that different applications warrant different approaches, depending on their requirements for data handling, performance, customization, and cost.

Muse Spark Open demonstrates that an advanced, customizable AI model can process information and generate responses across appropriately configured professional desktop computers, expanding the range of environments where such technology can operate. However, training massive AI models and operating them at scale still require substantial computing infrastructure, making data centers essential for many AI workloads. For organizations evaluating the next generation of AI applications, Meta's demonstration introduces greater flexibility in deciding how AI systems are deployed, managed, and integrated into their operations.

Meta's Muse Spark Open represents a meaningful expansion in how organizations can deploy advanced artificial intelligence. Its combination of competitive performance, downloadable model weights, and demonstrated operation on professional desktop computers gives businesses more flexibility in choosing the systems that power their AI applications. As AI agents take on increasingly complex responsibilities, the ability to select both the model and the infrastructure supporting it could become an important consideration in enterprise technology planning.

For business leaders, this introduces a more complicated purchasing decision. Organizations must evaluate whether the potential advantages of controlling their own AI infrastructure justify the costs and responsibilities involved. A company handling sensitive financial records, proprietary research, or confidential customer information may place greater value on managing where its AI processes data. Another business developing applications for customers may prioritize fast deployment, predictable expenses, and access to computing resources without maintaining specialized equipment. Both can use Muse Spark Open, but their requirements may lead to very different deployment decisions.

The model's performance also raises an important question about how organizations evaluate AI systems before trusting them with business operations. Muse Spark Open performed competitively in independent testing, yet its mistakes included overlooking hidden data patterns and incorrectly identifying valid information as faulty. These limitations become more consequential when AI agents can act on their conclusions, modify software, or interact with other business systems. Organizations considering those capabilities should test the model against their own workflows, measure the accuracy and consequences of its decisions, and establish appropriate oversight before allowing it to perform sensitive or irreversible actions. The ability to operate a model independently also places more responsibility for its security, reliability, and ongoing management on the organization using it.

As downloadable AI models become more capable and the hardware and software supporting them continue to evolve, organizations may gain additional ways to reduce dependence on individual AI service providers, customize systems for specialized work, and determine where their information is processed. Whether those benefits outweigh the costs of independent deployment will depend on the needs of each organization. Muse Spark Open provides another concrete example of how those choices are expanding, making the ability to evaluate, operate, and govern AI systems an increasingly important business capability. The companies that benefit most from greater AI independence will be those prepared to take responsibility for what their AI systems do.

Q: What is Meta's Muse Spark Open?
A: Muse Spark Open is an 816-billion-parameter open-weight AI model released by Meta on October 15, 2026. Designed for coding, complex reasoning, and AI-agent workflows, it activates approximately 42 billion parameters at a time rather than using its entire model for every calculation. Developers can download its trained weights, customize the model, and operate it on supported computing infrastructure under Meta's licensing terms.

Q: What can Muse Spark Open actually do?
A: Muse Spark Open can write and debug code, plan multistep tasks, use software tools, analyze information, and correct some errors during extended workflows. It accepts text, images, video, and audio as inputs, although its outputs are limited to text. Meta demonstrated the model making changes to an iOS application and operating professional video-editing software. These examples show its potential for AI-agent applications, but successful completion of complex tasks still requires appropriate testing and oversight.

Q: How is Muse Spark Open different from Meta's Muse Wonder?
A: Muse Spark Open is a downloadable model designed to give developers more flexibility in building and operating AI applications. Muse Wonder and Muse Wonder Full are Meta's more powerful proprietary models, accessible through Meta's hosted services. Muse Spark Open offers greater opportunities for independent deployment and customization, while the Wonder models provide more advanced capabilities through Meta-managed infrastructure.

Q: Where can I download or use Muse Spark Open?
A: Muse Spark Open's model weights are available through Meta's official repository on Hugging Face, subject to the repository's access requirements. Meta also provides developer documentation explaining how to configure and deploy the model. It is available through cloud model catalogs including Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Developers can also access hosted versions through providers including Together AI, Fireworks AI, and Baseten, which operate the model on their own infrastructure.

Q: Can regular people use Muse Spark Open, or is it just for developers?
A: Individuals can access Muse Spark Open under its licensing and access conditions, but the release is primarily designed for developers and organizations building AI applications. Running the model independently requires substantial computing resources and technical setup. Third-party hosted services provide another way to access the model without purchasing specialized hardware, although the availability of simple, consumer-friendly interfaces depends on the provider.

Q: Do I need Apple Mac Studios to run Muse Spark Open?
A: No. Meta and Apple demonstrated Muse Spark Open operating across two Mac Studios with 512 GB of unified memory each, but Apple hardware is only one supported option. Meta also supports deployment on NVIDIA, AMD, and Intel hardware. Additional compressed versions can reduce memory requirements, including a version described in Meta's documentation that fits into approximately 300 GB of combined memory on a single workstation. Running the model locally still requires substantial computing resources.

Q: Is Muse Spark Open free to use, including for businesses?
A: Meta makes Muse Spark Open's weights available without an upfront model licensing fee, and its license permits many commercial uses. However, users must comply with the license and usage policy, including required acknowledgments. Organizations offering AI models as a service must enter into a separate agreement with Meta if their combined gross revenue with affiliates exceeds $20 million during any consecutive 12-month period. The provision generally excludes internal business use and products that incorporate the model as a feature rather than offering AI models as a service.

Q: How much does it cost to run Muse Spark Open?
A: The cost depends on how the model is deployed. Apple's M5 Ultra Mac Studio starts at $5,499 in the United States, but the 512 GB configuration used in Meta's two-computer demonstration costs substantially more than the base model. Organizations running Muse Spark Open themselves must also consider infrastructure, electricity, maintenance, and technical support. Hosted providers offer alternatives without requiring hardware purchases, although usage charges vary. Downloading the model without an upfront licensing fee does not make it free to operate.

Q: How does Muse Spark Open compare with other AI models?
A: Muse Spark Open performs competitively against leading open-weight models, particularly in AI-agent tasks and data analysis, although it does not lead every benchmark. In independent testing by Plotly, it correctly answered 38 of 43 data-analytics questions, achieving an 88% score and tying two other open-weight models for the highest result on that leaderboard at the time. Meta's proprietary Muse Spark 1.3 and Muse Wonder scored higher, at 93% and 95%, respectively. Muse Spark Open also missed hidden data patterns and incorrectly flagged valid information as faulty, demonstrating limitations that matter for business applications.

Q: Is Muse Spark Open safe for sensitive business data?
A: Running Muse Spark Open on an organization's own infrastructure can allow prompts and proprietary information to remain within its controlled computing environment. However, that depends on how the surrounding applications and connected software tools handle information. Local deployment does not automatically guarantee privacy or security. Meta provided additional safety training for the downloadable model, but organizations remain responsible for implementing safeguards, testing reliability, controlling permissions, and establishing human oversight for consequential actions.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.