Meta has released Muse Spark Open, a new open-weight artificial intelligence model designed for complex software engineering, coding and autonomous AI tasks. Announced on October 15, 2026, the model contains 816 billion parameters, activates approximately 42 billion parameters per token and supports a context window of up to one million tokens. Meta says the model is designed to handle long-running workflows in which an AI system plans a task, uses tools, checks its work and corrects mistakes.
The release expands Meta’s efforts to compete in the increasingly competitive open-weight AI market, where developers can download model weights and customise how a system operates. However, Muse Spark Open is not fully open-source software under the conventional definition: Meta distributes it under the Muse Model License, which includes restrictions on certain commercial uses. The model is available through several AI infrastructure providers and model repositories, while local deployment requires substantial computing resources, including a configuration involving two Apple Mac Studio computers equipped with M5 Ultra chips.
Key Takeaways
- New model: Meta launched Muse Spark Open on October 15, 2026.
- Model size: It has 816 billion total parameters, with approximately 42 billion active per token.
- Context window: The model supports up to one million tokens, allowing it to process large amounts of contextual information.
- Main uses: Coding, complex software engineering, multimodal tasks and autonomous AI agents.
- Local deployment: Meta says it can run on two Mac Studio computers with M5 Ultra chips using Apple’s MLX framework.
- Commercial restrictions: The Muse Model License allows many uses but includes conditions for certain commercial applications.
- Availability: Developers can access the weights through Hugging Face and use supported inference platforms.
What Is Meta’s Muse Spark Open AI Model?
Muse Spark Open is an open-weight model developed by Meta Superintelligence Labs. Its architecture is based on a sparse mixture-of-experts approach, which allows the system to activate only part of its total parameter capacity for a given token.
In a traditional dense model, most of the model’s parameters participate in processing each token. A mixture-of-experts model routes inputs through selected components, potentially making large-scale models more computationally efficient than a dense architecture of comparable total size.
Muse Spark Open contains 816 billion parameters in total but activates approximately 42 billion per token. This does not mean the model is small or inexpensive to operate: its full weights still need to be stored across suitable hardware, and actual performance depends on memory bandwidth, inference software, quantisation and workload.
Meta is positioning the model for tasks that require more than a single response. Instead of simply answering a coding question, an AI agent can inspect a codebase, plan changes, modify files, run tests and revise its work when something fails.
This approach is increasingly important as developers move from AI-assisted autocomplete toward systems that can complete larger portions of a software project.
Muse Spark Open Specifications
| Specification | Details |
|---|---|
| Model name | Muse Spark Open |
| Developer | Meta Superintelligence Labs |
| Total parameters | 816 billion |
| Active parameters per token | Approximately 42 billion |
| Context window | Up to 1 million tokens |
| Architecture | Sparse mixture-of-experts |
| Input capabilities | Text, image, video and audio |
| Primary applications | Coding, agentic workflows and complex reasoning |
| Licence | Muse Model License |
| Local hardware example | Two Apple Mac Studio systems with M5 Ultra chips |
Specifications are based on Meta’s launch documentation. Actual deployment requirements vary by model precision, inference engine and workload.
Why Meta Is Targeting Coding and Agentic AI
AI coding tools have evolved rapidly from systems that suggest a line of code to agents that can work across multiple files, interact with development environments and attempt to complete software tasks independently.
These workflows require several capabilities. The model must understand a user’s objective, maintain context across a large codebase, identify dependencies, decide which tools to use and evaluate whether the changes actually work.
A system that performs well on short coding questions may struggle with a task that involves modifying multiple files, running tests and correcting unexpected failures. Long-running agentic work therefore places different demands on a model than ordinary question-and-answer conversations.
Meta says Muse Spark Open is built for long-horizon tasks, complex software engineering and multi-step reasoning. These are company claims about the model’s capabilities, and developers should evaluate them against their own projects and independent benchmarks before choosing it for production use.
How an AI coding agent works
A coding agent typically follows a cycle:
- Understand the requested change and inspect the relevant files.
- Develop a plan for implementing the change.
- Edit code or invoke tools such as a terminal and test runner.
- Examine errors, test results and other feedback.
- Correct problems and verify the final result.
The model’s context window can help it retain information about a large project, while tool use allows it to interact with the software environment. However, a large context window does not guarantee that a model will correctly understand every file or make reliable changes.
Human review, testing and appropriate permissions remain important, particularly when agents can access repositories, credentials or production systems.
What Does the One-Million-Token Context Window Mean?
A context window determines how much information a model can consider within a single interaction or task, subject to the implementation and the available input and output budget.
Muse Spark Open’s one-million-token context window is designed to support workflows involving large codebases, extensive technical documents and long sequences of agent interactions.
For software developers, this could mean supplying more source code, documentation, test results and project instructions without repeatedly splitting the information into smaller chunks.
The capability may also help with tasks such as reviewing a large repository, tracing dependencies or comparing changes across multiple files.
However, context capacity should not be confused with perfect recall or reasoning. A model can have access to a large amount of information and still overlook relevant details, misunderstand dependencies or make incorrect decisions. Performance also depends on how the context is organised and how the model uses the information it receives.
Meta Says Muse Spark Open Can Run on Apple Mac Studio
One of the most unusual aspects of the release is Meta’s focus on local deployment using Apple hardware.
Meta says Muse Spark Open can run across two Mac Studio computers equipped with M5 Ultra chips using Apple’s MLX machine-learning framework. The configuration takes advantage of Apple’s unified memory architecture, which allows the integrated GPU to access a large pool of system memory.
Running a model locally can provide greater control over data and infrastructure. Developers may be able to experiment with private code, customise model behaviour and avoid sending every prompt to a third-party hosted service.
But the hardware requirements are substantial. Two high-end Mac Studio systems represent a significant investment, and local operation also involves power consumption, setup, memory management and performance tuning.
For organisations that do not want to purchase and operate their own hardware, hosted inference providers offer another route. These services run the model on their infrastructure and charge according to their pricing and usage arrangements.
The practical choice depends on data sensitivity, budget, latency, expected workload and the technical expertise available to maintain the deployment.
Where Developers Can Download or Run Muse Spark Open
Meta has made the model weights available through Hugging Face and several cloud model catalogues. It is also supported by multiple inference providers, allowing developers to experiment without necessarily building an entire deployment stack from scratch.
According to Meta’s release documentation, the model is available through platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Its weights are also listed through Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Developers can use supported inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX and llama.cpp, subject to the appropriate hardware, model format and configuration.
The model can also be customised using supported machine-learning frameworks. That flexibility is one of the principal advantages of open-weight releases: developers can adapt the system to specific applications instead of relying exclusively on a provider’s hosted interface.
However, running the model successfully still requires technical knowledge. Users need to understand hardware compatibility, memory requirements, inference configuration, security and the licence governing their intended use.
Open-Weight Does Not Mean Fully Open-Source
The licensing terms are an important consideration for businesses evaluating Muse Spark Open.
Open-weight models provide access to trained parameters, commonly called weights. Those parameters can be downloaded and used with compatible software, and developers may be able to modify or fine-tune the model.
Fully open-source AI can involve broader access to the components needed to study, reproduce and modify the system, subject to the relevant licence. An open-weight release does not automatically provide all training data, complete training code or every detail of the development process.
Meta distributes Muse Spark Open under its Muse Model License rather than an unrestricted permissive licence such as Apache 2.0. The licence allows many developers to use the model, but includes conditions on certain commercial applications and high-scale use.
Some reporting has highlighted a $20 million threshold associated with commercial-use terms. Businesses should consult the actual licence for the precise definition, scope, calculation period and obligations before deploying the model commercially.
The distinction matters because companies need to understand their legal obligations before integrating a model into a paid product, offering it as a service or distributing a modified version. Download availability alone does not establish that every possible use is permitted without additional conditions.
How Muse Spark Open Competes With Other AI Models
The open-weight AI market has become increasingly competitive, with developers comparing models on coding, reasoning, tool use, price, hardware requirements and licence flexibility.
Chinese developers including Alibaba, Moonshot AI and Z.ai have built prominent model families that are available in open-weight forms. Meta is positioning Muse Spark Open as a leading US alternative, particularly for coding and agentic workloads.
| Competitive factor | Why it matters |
|---|---|
| Coding performance | Determines how reliably the model can modify, debug and test software |
| Agentic capabilities | Affects its ability to complete multi-step tasks using tools |
| Context window | Helps when working with large repositories and long documents |
| Hardware efficiency | Influences local deployment costs and response speed |
| Licence conditions | Determines which commercial uses are permitted |
| Multimodal support | Enables applications involving images, video and audio as well as text |
Meta’s claims about its position in the market should be distinguished from independent evaluations. Results depend on the benchmark, testing methodology, model configuration and comparison set. Developers should test the model against alternatives using representative workloads rather than relying on a single headline claim.
What the Launch Means for AI Developers and Businesses
Muse Spark Open could be useful for teams seeking greater control over AI infrastructure and model behaviour. Its open weights, broad inference-engine support and agentic focus give developers several possible deployment paths.
A software company could evaluate it for code review, bug fixing, test generation or internal developer tools. A business with strict data-handling requirements could investigate local deployment, provided it has sufficient hardware and can satisfy the licence.
The model’s size also creates barriers. Smaller teams may find hosted inference more practical than running it locally. Others may prefer smaller models that require less memory and offer lower latency for simpler tasks.
Businesses must also consider security. Coding agents can modify files, execute commands and interact with external systems. Their permissions should be limited, changes reviewed and sensitive operations tested in controlled environments.
The model’s usefulness will ultimately depend on the quality of its outputs, the cost of running it and the reliability of its agentic workflows—not just its parameter count.
The Bigger Picture
Muse Spark Open highlights Meta’s continued investment in developer-controlled AI systems as competition shifts toward models that can perform complex tasks rather than merely generate text. Its combination of a large total parameter count, a one-million-token context window, multimodal inputs and agentic capabilities is aimed at developers building increasingly autonomous software tools.
The release also demonstrates that openness is not a simple yes-or-no category. Downloadable weights provide flexibility, but licensing conditions and demanding hardware can limit how easily organisations adopt a model. For developers, the most important decision will be whether Muse Spark Open delivers useful results at an acceptable cost under the terms and infrastructure available to them.
Looking Ahead
Developers will be able to assess Muse Spark Open through hosted inference platforms, downloadable weights and supported local deployment tools. Its performance in real software projects, independent benchmark evaluations and the practical cost of running it will help determine whether it becomes a widely adopted option for coding and agentic applications.
The broader competition will depend on more than raw model size. Efficiency, reliability, security, customisation and licence flexibility will influence which systems businesses choose. Muse Spark Open gives Meta another entry in that competition, while leaving developers responsible for testing the model and verifying that it fits their technical and commercial requirements.
Frequently Asked Questions
1. What is Meta Muse Spark Open?
Muse Spark Open is Meta’s open-weight AI model designed for coding, complex software engineering, multimodal processing and autonomous agentic tasks.
2. How large is Muse Spark Open?
It contains 816 billion total parameters, activates approximately 42 billion parameters per token and supports a context window of up to one million tokens.
3. Is Muse Spark Open free to download?
Meta makes the model weights available to download, but its Muse Model License includes conditions on certain commercial uses. Users should review the licence before deployment.
4. Can Muse Spark Open run on a personal computer?
It requires substantial computing resources. Meta describes a local setup using two Apple Mac Studio computers with M5 Ultra chips. It is not a typical lightweight model for an ordinary laptop.
5. What can developers build with Muse Spark Open?
Potential applications include coding agents, software debugging tools, repository analysis, automated testing and other systems that require multi-step reasoning and tool use.
6. Is Muse Spark Open fully open-source?
It is an open-weight model, which means its trained weights are available. That does not automatically make it fully open-source software, and its licence imposes conditions that developers must review.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
