TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Meta says Muse Spark Open is available as open weights and through hosted inference providers. The multimodal model has 816 billion total parameters, 42 billion active parameters and a 1 million-token context window; Meta highlights coding and agentic tasks, including a desktop setup using two Apple M5 Ultra Mac Studios. Its license has limited commercial terms, and details of its safety evaluations are not yet fully published.
Meta has released the weights for Muse Spark Open, an open-weight, natively multimodal model designed for coding and agentic tasks. Meta says the model has 816 billion total parameters, with 42 billion active at a time, and a 1 million-token context window; developers can download it or access hosted versions through listed inference providers.
Meta describes Muse Spark Open as its most capable US open-weight model to date, a company characterization rather than an independently established ranking. The company says the model is built for long-horizon tasks, complex software engineering and multi-step reasoning. The source points to a separate evaluation methodology report but does not include benchmark scores, so its performance claims cannot be assessed from the announcement alone.
The weights are listed for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, according to Meta. The company names SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth among supported inference engines, with optimizations for several chip platforms.
Meta also says the model can run locally across two Apple M5 Ultra Mac Studios using the MLX framework. It cites workflows run on-device with its Muse Code tool, including changes to the Ice Cubes iOS app and editing video in DaVinci Resolve. These are company-reported demonstrations; the announcement does not give independent testing results or a full account of the hardware and runtime conditions.
More Access, With Hardware Limits
Releasing model weights gives developers more control than access through a hosted chatbot alone: they can run the model on supported infrastructure, adapt it, and integrate it into their own software. That could be relevant to teams building coding agents and multi-step workflows, particularly if the model performs as Meta says on complex tasks.
Local use is a prominent part of the announcement, but it is not equivalent to running the model on an ordinary desktop. Meta’s demonstration uses two high-memory Mac Studio systems, and its account depends on Apple’s unified memory and the MLX software stack. The release expands options for developers, while the hardware requirements and limited commercial terms may shape who can practically use it.
As an affiliate, we earn on qualifying purchases.
How Meta Is Distributing the Model
The release combines downloadable weights with hosted inference and support for multiple inference engines. This offers developers different ways to try the model: use a third-party platform, download the files from a listed repository, or run it on compatible hardware. Meta says further customization is possible using PyTorch and TorchTitan.
The model is distributed under the Muse Model License, which Meta describes as allowing many developers to use it for free while imposing limited commercial terms. The announcement does not set out those terms in full. Developers considering commercial deployment will need to review the license itself rather than assume the weights are available under an unrestricted open-source license.
““The most capable US open-weight model to date.””
— Meta AI
As an affiliate, we earn on qualifying purchases.
Benchmarks and Safety Details Pending
The announcement does not include the numerical benchmark results or the full evaluation methodology needed to compare Muse Spark Open with other models. Meta says further evaluation information is available in a methodology report, but the supplied release does not reproduce those results.
Safety claims are also based on Meta’s own testing under its Superintelligence Scaling Framework. The company says the model met its required release thresholds and received additional safety post-training, but says a full evaluation report is forthcoming. The announcement does not provide independent validation, detailed risk-category scores, or a release date for that report. Practical performance, hardware costs and the precise scope of the license also remain questions for developers.
As an affiliate, we earn on qualifying purchases.
What Developers Can Test Next
Developers can access the weights through the listed repositories, try hosted inference from Meta’s named partners, or use supported runtimes for local deployment. Those tests will help establish how the model performs on real workloads, how much hardware and configuration it needs, and whether it is suitable for particular software or agent workflows.
Meta says a detailed evaluation report will be available soon, without giving a date in the announcement. The next useful evidence will include that report, independent benchmark comparisons and practical accounts of deployment costs and results. Developers planning commercial use should also check the Muse Model License for its exact restrictions.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is Meta’s open-weight, natively multimodal model, described by the company as designed for coding, agentic work and multi-step reasoning. Meta lists 816 billion total parameters, 42 billion active parameters and a 1 million-token context window.
Where can developers get the model?
Meta lists downloads through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. It also names several hosted inference providers, including Baseten, Groq and Together AI.
Can it run on a desktop?
Meta says it can run locally with Apple MLX across two Apple M5 Ultra Mac Studios. That is a specific high-end setup, not evidence that the model will run on typical consumer computers.
Is Muse Spark Open unrestricted for commercial use?
No unrestricted commercial license is stated. Meta says the weights use the Muse Model License, which has limited commercial terms and allows many developers to use the model for free. Users should review the license for exact conditions.
Are its benchmark and safety claims independently verified?
The announcement presents Meta’s own claims and says more evaluation material is available or forthcoming. It does not include full benchmark results or independent verification; Meta says a full evaluation report will be published later.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
