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 today, with model weights for download and hosted access through several platforms. The company describes it as an 816-billion-parameter multimodal model that can run locally across two Apple M5 Ultra Mac Studios; its license has limited commercial terms.
Meta has released the weights for Muse Spark Open, an 816-billion-parameter open-weight model built for coding, multimodal input and multi-step agentic work. The company says developers can download the weights or access the model through hosted inference providers, and that it can run locally across two Apple M5 Ultra Mac Studios.
Meta describes Muse Spark Open as a natively multimodal model with 816 billion total parameters, of which 42 billion are active, and a 1-million-token context window. The company says it is designed for long-running tasks, complex software engineering and multi-step reasoning. Those descriptions are Meta’s characterization; the release announcement does not provide independent comparative testing results in the supplied material.
The weights are listed for download through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is offered through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta says the model supports inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with support for hardware from several manufacturers.
Meta says the model’s Muse Model License allows many developers to use it free, but includes limited commercial terms. The announcement also says developers can customize it with PyTorch’s TorchTitan. Availability through the named services does not mean every provider offers identical pricing, performance or access conditions.
Local Access to a Large Model
The release gives developers another option between using a hosted model and running model weights on their own hardware. Downloadable weights and multiple inference engines can give teams more control over deployment and customization, while hosted providers may reduce the need to operate specialized infrastructure themselves. The license’s limited commercial terms may affect whether and how some businesses can use it.
Meta’s desktop claim is notable because the model’s stated size makes local operation a substantial hardware task. The company says two M5 Ultra Mac Studios, using Apple’s unified-memory architecture, can hold the model. That is a vendor-reported configuration, not evidence that typical desktop systems can run it or that local performance will match hosted services. The announcement does not give a complete comparison of cost, speed or power use across deployment options.
As an affiliate, we earn on qualifying purchases.
How Meta Says It Runs Locally
Meta says it optimized Muse Spark Open for Apple Silicon using the MLX framework. According to the report, each M5 Ultra Mac Studio can provide up to 512 GB of unified memory, allowing two systems to hold the model. Meta describes their combined capacity as a terabyte of unified memory and says the setup can run agentic workflows on-device.
As examples, the company says it used its Muse Code system with Muse Spark Open to make and build changes to Ice Cubes, an open-source Mastodon client, and to operate DaVinci Resolve. In the latter demonstration, Meta says the model reviewed footage frame by frame to identify moments and create vertical crops. These are company-reported demonstrations; the announcement does not provide independent validation or enough detail to assess how reliably the workflows perform across other projects.
For graphics processors, Meta says its release includes custom, auto-tunable Helion kernels for SGLang. It reports kernel-level speedups of 1.03 to 3.95 times on NVIDIA and AMD GPUs, and a 3% end-to-end speedup on AMD MI350. The kernel figures apply to particular operations, while the end-to-end result is a separate measure; the supplied report does not specify a comparison baseline or test conditions for these numbers.
““The most capable US open-weight model to date.””
— Meta
Performance and License Details
The announcement does not give a calendar date, complete benchmark tables or enough methodology to independently assess the model’s stated capabilities. Meta refers readers to an evaluation methodology report and says a full evaluation report will be available soon. It also says the model met release thresholds under its Meta Superintelligence Scaling Framework and received additional safety post-training, but the supplied material does not contain the detailed results or define all relevant test conditions.
It remains unclear how the model performs against other models on independent evaluations, what hardware and software settings produced each reported speed result, and what deployment costs users should expect. The license is described as allowing many uses for free subject to limited commercial terms; developers should consult the actual license for the restrictions that apply to their projects.
Evaluation Reports and Adoption
Meta says it plans to publish a full evaluation report and points to the model card for safety and release information. Those materials may provide more detail on evaluation methods, risk testing and the basis for the company’s performance claims. Developers can access the weights through the listed repositories and try hosted versions through partner platforms, subject to each provider’s availability and terms.
The next practical test will be how the model performs in independent use across coding, multimodal and agentic workflows, and whether its local configuration is feasible for teams beyond the hardware setup Meta describes. Pricing, commercial permissions and provider-specific performance will need to be checked with the relevant services and the license.
Key Questions
What is Muse Spark Open?
Meta describes it as an open-weight, natively multimodal model with 816 billion total parameters, 42 billion active parameters and a 1-million-token context window, aimed at coding and agentic tasks.
Where can developers get the model?
Meta lists downloads through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted access is listed through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Can it run on a desktop computer?
Meta says it can run locally across two Apple M5 Ultra Mac Studios using MLX. The announcement does not establish that it will run on ordinary desktop hardware or provide comparable performance on smaller systems.
Is Muse Spark Open free for commercial use?
Meta says its Muse Model License allows many developers to use the model for free, but includes limited commercial terms. The applicable permissions depend on the license itself, which users should review before deployment.
How have the model’s safety claims been evaluated?
Meta says it conducted evaluations under its own Superintelligence Scaling Framework, added safety post-training and met its release thresholds. The detailed evaluation report was not included in the supplied announcement and is expected later, according to the company.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
