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 now as an open-weight model for coding and agentic tasks. It has 816 billion total parameters, a 1 million-token context window and a setup designed to run across two Apple M5 Ultra Mac Studios; Meta’s capability and safety descriptions are company claims.
Meta has released the Muse Spark Open weights, making its 816-billion-parameter, natively multimodal model available for download and hosted use. The company says it is built for coding and agentic work and can run locally across two Apple M5 Ultra Mac Studios, a release that gives developers another option for running a large model outside a single cloud service.
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 capable in long-horizon tasks, complex software engineering and multi-step reasoning. Those performance descriptions are the company’s claims; its post points to a separate evaluation methodology report, but the supplied announcement does not give benchmark scores or comparisons.
The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, according to Meta. The company lists support for engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and Nvidia hardware.
For local use, Meta says the model was optimized with Apple’s MLX framework for two M5 Ultra Mac Studios. Each system can provide up to 512 GB of unified memory, according to the announcement, allowing the two-machine setup to hold the model. Meta also reports custom Helion GPU kernels in SGLang, with 1.03–3.95× kernel-level speedups on Nvidia and AMD GPUs and a 3% end-to-end speedup on AMD MI350. The kernel figures are not the same as an overall model-performance comparison.
Local Access to a Large Model
The release matters because it pairs downloadable weights with several ways to run the model: hosted inference, different inference engines and a local Apple Silicon configuration. Developers can choose among providers or adapt the model for their own workflows, rather than relying only on Meta’s own interface. How practical those options are will depend on hardware costs, deployment requirements and measured performance, none of which are fully detailed in the announcement.
Meta says it tested agentic workflows on-device using Muse Code, including making changes to an open-source iOS app and operating DaVinci Resolve to review footage and select vertical crops. These examples describe the company’s tests, not independent verification or a guarantee that the model will perform similarly across other projects. The release also raises the usual open-weight questions about the license, safety practices and what users can do after downloading the model.
As an affiliate, we earn on qualifying purchases.
Release Channels and License
Meta is releasing the weights under its Muse Model License, which it says has limited commercial terms and allows many developers to use the model for free. The announcement does not reproduce the full license conditions, so developers should review the license itself before using the model in a commercial product or service.
The company says it added safety post-training for the open-weight release and assessed the model under its Meta Superintelligence Scaling Framework. Meta reports that Muse Spark Open met the framework’s release thresholds across the severe-risk categories it measured. The full evaluation report was described as forthcoming; the model card is the source Meta identifies for detailed results.
““The most capable US open-weight model to date.””
— Meta
large language model inference hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and Safety Detail
The supplied announcement does not provide benchmark scores, independent evaluations or a detailed comparison with other open-weight models. Meta’s claim that this is the most capable US open-weight model is therefore not substantiated with comparative figures in the source material. It is also not clear what performance users should expect across the listed hardware and inference engines, or what the complete cost and setup requirements are for running the model locally.
Meta says detailed safety results are in the model card and that a full evaluation report will be available soon, but the source does not give a publication date. The exact limits imposed by the Muse Model License are also not set out in the announcement. Those points may affect how organizations assess the model for deployment.
As an affiliate, we earn on qualifying purchases.
Evaluation and Deployment Details
Developers can access the weights and hosted options now, according to Meta, and can use PyTorch’s TorchTitan to further tune the model. The company’s next stated information release is a full evaluation report; it did not provide a date. Readers assessing the model should compare that report and the model card with independent tests, review the license terms, and check hardware-specific results as implementations become available.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is Meta’s open-weight, natively multimodal model, with 816 billion total parameters, 42 billion active parameters and a 1 million-token context window, according to the company.
Where can developers get the model?
Meta says weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference is also listed through seven partner platforms.
Can it run on a desktop computer?
Meta says it can run locally using two Apple M5 Ultra Mac Studios with the MLX framework. The announcement does not establish that the model can run on a typical desktop or a single Mac.
What license applies?
The weights are released under the Muse Model License, which Meta describes as having limited commercial terms. Users should consult the license text for the precise permissions and restrictions.
Are the model’s safety and performance claims independently verified?
The supplied source reports Meta’s own evaluations and claims, but does not include independent verification or full benchmark results. Meta says a detailed evaluation report will be available soon.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
