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 has released the weights for Muse Spark Open, a multimodal model it describes as its most capable U.S. open-weight model to date. The company says it can run on two Apple M5 Ultra Mac Studios and is available through several inference providers and model repositories; its safety evaluations and performance claims are Meta’s own.
Meta has released the weights for Muse Spark Open, a multimodal model the company calls its most capable U.S. open-weight model to date, making it available for download and through multiple hosted inference providers. Meta says the model has 816 billion total parameters, with 42 billion active, and can run on a desktop setup using two Apple M5 Ultra Mac Studios, a claim that could broaden access to high-capacity models beyond centralized cloud services.
Meta describes Muse Spark Open as built for coding and agentic work, including long-running tasks, complex software engineering and multi-step reasoning. It reports a one-million-token context window and says developers can customize the model using PyTorch. Those descriptions come from Meta; the announcement does not provide independent test results in the supplied material.
The weights are available 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. Meta says the model supports inference 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 Apple hardware, Meta says its MLX optimization allows two M5 Ultra Mac Studios to hold the model using their combined unified memory. The company also reports custom, auto-tunable Helion GPU kernels in SGLang, with 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. These are vendor-reported results; the announcement points to an evaluation methodology report for more detail.
Open Weights Bring Models to Desktops
The release gives developers the option to download and adapt the model rather than relying only on a hosted product. That can support experiments with local inference, custom tuning and different hardware, while the listed hosting partners offer a route for users who do not have the equipment to run it themselves.
Meta’s desktop demonstration is notable because it describes running the model across two workstations rather than a data-center deployment. The company says it used Muse Code with Muse Spark Open to make changes to the open-source iOS app Ice Cubes, including bug fixes and new features, and to operate DaVinci Resolve for video review and vertical cropping. Those examples illustrate Meta’s intended use cases, but they do not establish how reliably the model performs across other projects or users’ machines.
Open weights can also make scrutiny and modification possible beyond the model developer, while increasing the need for users and deployers to evaluate behavior in their own settings. Meta says it added safety post-training and tested severe-risk categories under its internal framework. The release therefore combines wider developer access with questions about licensing, deployment safeguards and the evidence behind performance claims.
As an affiliate, we earn on qualifying purchases.
How Meta Says the Desktop Setup Works
Meta says the Apple configuration uses the unified memory architecture of M5 Ultra Mac Studios. According to the company, each system can provide up to 512 GB of memory to its integrated GPU, allowing two machines to hold the model together. Meta also cites memory bandwidth of 1.2 terabytes per second for the chip and says the systems can act as a single pool of memory for inference.
The release is not limited to Apple systems. Meta lists support across several inference engines and chip makers, and points developers to PyTorch TorchTitan for further tuning. The weights are licensed under the Muse Model License, which Meta says includes limited commercial terms and allows many developers to use the model for free. The announcement does not spell out every permitted use or restriction; developers will need to consult the license itself.
“Today, we’re releasing the weights for Muse Spark Open, the most capable US open-weight model to date.”
— Meta, in its Muse Spark Open announcement
As an affiliate, we earn on qualifying purchases.
Independent Testing and License Terms
The announcement does not include the full benchmark results, independent evaluations or the complete safety findings. Meta says its assessment followed the Meta Superintelligence Scaling Framework, that the model met its release thresholds, and that a full evaluation report would be available soon. The supplied material does not establish how the model compares with other systems under independently run tests.
It is also not clear from the announcement how performance varies across the supported hardware and inference engines, or what practical limits users may encounter when running the model locally. Meta’s stated speedups refer to particular kernel and hardware measurements, not a general guarantee of faster overall performance. The license’s limited commercial terms may also affect particular uses, so developers should review its text before deploying the model commercially.
high-performance desktop GPU for AI
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Evaluation Reports and Developer Testing
Meta says a full evaluation report is coming, with additional methodology information referenced in its announcement. That report should provide more detail on benchmarks and safety testing. Developers can access the weights through the listed repositories, use hosted inference partners, or try the Apple MLX setup described by Meta.
The next useful evidence will come from independent testing of model quality, speed, safety and resource requirements across different setups. Meta has invited developers to share projects in the Hugging Face Forums, but the announcement does not give a date for the evaluation report or identify a separate release schedule for future updates.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did Meta announce?
Meta announced that the Muse Spark Open weights are available for download, alongside access through hosted inference providers. The company says the model is multimodal and intended for coding and agentic tasks.
Can Muse Spark Open run on a desktop?
Meta says it can run on two Apple M5 Ultra Mac Studios using the MLX framework and their combined unified memory. The announcement describes Meta’s configuration; independent confirmation and performance comparisons were not provided in the supplied material.
Where can developers access the model?
Meta lists Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models for weights. It also names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as hosted inference providers.
Is the model free for commercial use?
Meta says the Muse Model License has limited commercial terms and permits many developers to use the model for free. Specific commercial uses may be subject to restrictions, so the license should be checked directly.
Have the safety claims been independently verified?
The announcement reports Meta’s internal safety evaluation and says the model met the company’s release thresholds. A full report is described as forthcoming, and the supplied material does not include independent safety testing.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
