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 weights are available now through several model platforms, with hosted inference also offered by listed partners. The company describes it as an 816-billion-parameter, natively multimodal model that can run on two Apple M5 Ultra Mac Studios; its safety findings and performance claims are based on Meta’s own evaluations.
Meta has released the Muse Spark Open model weights, making its 816-billion-parameter model available for download and through hosted inference services. The company says the natively multimodal model is built for coding and agentic work and can run locally across two Apple M5 Ultra Mac Studios, a setup that brings high-capacity model use onto desktop hardware for developers with suitable equipment.
Muse Spark Open has 816 billion total parameters, of which Meta says 42 billion are active, and a 1-million-token context window. Meta describes it as suited to long-running tasks, complex software engineering and multi-step reasoning. Those are the company’s characterizations; the announcement points readers to a separate evaluation methodology report for details on its benchmarks.
Weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta also lists hosted inference through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The model supports several inference engines, including SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth, with stated optimizations for AMD, Apple, Intel and Nvidia hardware.
Meta says users can customize the model with PyTorch and TorchTitan. The weights use the Muse Model License, which the company says has limited commercial terms and permits many developers to use the model for free. Specific licensing conditions should be checked in the license itself before commercial deployment.
Local Access Needs Two Mac Studios
The release gives developers a choice between hosted inference and running the model on hardware they control. Local deployment can matter to teams working with sensitive code or data, though the announcement does not quantify privacy, operating costs or practical performance relative to hosted services.
The desktop claim also comes with a substantial hardware requirement: Meta’s described configuration uses two M5 Ultra Mac Studios. The company says each machine can provide up to 512 GB of unified memory and that the systems together can hold the model. This is not a claim that the model will run on a typical personal computer; access depends on specialized, high-memory equipment.
Meta reports that its custom Helion GPU kernels in SGLang delivered 1.03 to 3.95 times kernel-level speedups on Nvidia and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. These figures refer to the company’s stated tests and should not be treated as independent comparisons across all workloads or systems.
Products Worth Considering
Apple Watch Ultra 2 GPS + Cellular, 49mm Natural Titanium Case with Natural Titanium Milanese Loop, Large
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID – Midnight
Bose Lifestyle Ultra Wireless Home Speaker
As an affiliate, we earn on qualifying purchases.
Meta’s Open-Weight Release
Meta says the release is part of its effort to make advanced models available beyond its own products. Open weights allow developers to download and run model parameters, subject to the applicable license; they do not, by themselves, mean every element of a model’s training data or development process is public.
For its local-use demonstration, Meta says it ran agentic workflows through Muse Code on the two-Mac setup. Examples included making and building changes to Ice Cubes, an open-source Mastodon client, and using DaVinci Resolve to review footage frame by frame and select moments for vertical cropping. These are company-reported demonstrations, not independent assessments of reliability across projects.
The announcement also describes additional safety post-training for the open-weight release. Meta says the model met release thresholds under its Superintelligence Scaling Framework, which evaluates severe risks. It directs readers to the model card for detailed results and says a fuller evaluation report will be available later.
““The most capable US open-weight model to date.””
— Meta
Independent Results Still Pending
The announcement does not specify a calendar release date, provide the full benchmark tables, or include independent testing of its performance claims. Meta says an evaluation methodology report is available and a full safety evaluation report will follow; the scope and timing of that report are not stated in the supplied material.
It is also unclear how the model performs across different hardware configurations, how much memory or time specific tasks require, and what the cost of hosted access will be. Meta’s safety statement reflects its own framework and release thresholds; the source does not supply enough detail to independently assess the underlying results.
Developers Test the Released Weights
Developers can download the weights from the listed model repositories or try the model through Meta’s named inference partners. They can also use supported inference engines and, for Apple silicon, the MLX adapter described by Meta. Practical reports from users may clarify how well the model’s claimed capabilities translate across hardware and real workloads.
Meta says a full evaluation report will be published soon and invites developers to share projects on the Hugging Face forums. The next useful milestones are the release of that report, more detailed benchmark results and independent testing of both local and hosted deployments.
Key Questions
What is Muse Spark Open?
It is an open-weight model released by Meta. Meta describes it as natively multimodal, with 816 billion total parameters, 42 billion active parameters and a 1-million-token context window.
Where can developers access it?
Meta lists downloads through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, plus hosted inference through seven named service providers.
Can it run on a desktop computer?
Meta says it can run locally using two Apple M5 Ultra Mac Studios with MLX. The announcement does not say that a standard desktop or a single Mac can run the full model.
Is the model free for commercial use?
Meta says the Muse Model License allows many developers to use the model for free but includes limited commercial terms. Users should review the license to determine whether a specific use is permitted.
Have the safety and benchmark claims been independently verified?
The supplied announcement reports Meta’s own evaluations. It does not include independent verification; Meta points to its model card and says a full evaluation report is forthcoming.
Source: hn
