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, with hosted access through several inference providers and downloads through major cloud and model platforms. The company says it can run locally across two Apple M5 Ultra Mac Studios and was evaluated under Meta’s safety framework; independent performance and safety results are not included in the supplied announcement.
Meta has released the weights for Muse Spark Open, an 816-billion-parameter multimodal model aimed at coding and agentic tasks, and says it can run on a desktop setup built from two Apple M5 Ultra Mac Studios. The release gives developers access through model downloads and hosted inference services, while Meta describes the weights as open under a license with limited commercial terms.
Meta reports that Muse Spark Open has 816 billion total parameters, of which 42 billion are active, and a one-million-token context window. The company says the model is built for long-running tasks, complex software engineering and multi-step reasoning. Those descriptions are Meta’s claims; the announcement points readers to a separate evaluation methodology report but does not provide benchmark results in the material supplied here.
The 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. It says the model supports inference engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for several chip platforms.
Meta says its MLX implementation can use the unified memory of two M5 Ultra Mac Studios to hold the full model. The company describes the pair as providing a terabyte of unified memory and says it tested on-device workflows involving changes to an open-source iOS app and video editing in DaVinci Resolve. The announcement does not independently verify those demonstrations or specify how long the tasks took, what human involvement they required, or the setup’s total cost.
Local Access to a Large Model
The release broadens access to a model Meta positions for software development and multi-step computer tasks. Developers can download weights, adapt the model, or use listed hosted services rather than relying solely on Meta’s own interface. Local execution may also appeal to teams that want more control over where model workloads run, although the described desktop configuration requires two high-end workstations rather than an ordinary consumer computer.
The scale of the model makes its hardware and licensing terms relevant to practical adoption. Meta says its license allows many developers to use it for free but includes limited commercial terms. That is not the same as unrestricted use, and organizations will need to review the license for their intended applications. The company’s stated kernel and hardware optimizations may improve access on some systems, but reported speedups do not by themselves establish overall task quality or cost effectiveness.
As an affiliate, we earn on qualifying purchases.
How Meta Says It Runs
Meta presents Muse Spark Open as a way to run an advanced model both through cloud inference providers and on locally controlled hardware. Its desktop claim depends on Apple’s unified-memory architecture and the latest MLX framework, with two M5 Ultra Mac Studios working together. Meta says the M5 Ultra can make up to 512 GB of memory available to its integrated GPU and cites memory bandwidth of 1.2 terabytes per second.
For GPU execution beyond Apple hardware, Meta says the release includes custom, auto-tunable Helion kernels for SGLang. It reports kernel-level speedups ranging from 1.03 to 3.95 times on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. The first figures are specifically kernel-level results, not a general statement that every workload runs several times faster. Meta also says developers can customize the model with PyTorch’s TorchTitan.
““We’re releasing the weights for Muse Spark Open, the most capable US open-weight model to date.””
— Meta AI
large language model inference server
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmarks and License Details
The announcement does not provide the benchmark scores or full evaluation methodology needed to assess the model’s coding, reasoning or agentic performance against alternatives. Meta says a methodology report contains more detail and that a full safety evaluation report will be available soon. The company also says the model met release thresholds in its own Meta Superintelligence Scaling Framework, but the announcement does not provide the underlying results for review.
Other practical details remain unspecified, including the model’s performance across different hardware configurations, the time and cost required for the showcased workflows, and the extent of human direction in those demonstrations. The exact limits in the Muse Model License are also not set out in the supplied announcement, so developers should consult the license itself before commercial deployment.
As an affiliate, we earn on qualifying purchases.
Reports and Developer Testing
Developers can access the weights through the listed repositories and cloud model platforms, or try hosted inference through Meta’s named partners. Those routes should let users test the model against their own coding and agentic workloads and determine how it performs on available hardware.
Meta says a full evaluation report is forthcoming. Its publication would provide more detail on the safety assessments and release thresholds described by the company. Further independent testing and clearer reporting on workload-level performance, hardware requirements and licensing will help establish how the model compares with other available systems.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Muse Spark Open?
It is an open-weight multimodal model from Meta, with 816 billion total parameters, 42 billion active parameters and a one-million-token context window, according to the company.
Where can developers get the model?
Meta lists downloads through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference 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 using two Apple M5 Ultra Mac Studios with MLX. The announcement does not claim that the full model will run on a typical desktop or provide a lower-cost hardware configuration.
Is Muse Spark Open unrestricted for commercial use?
No such conclusion can be drawn from the announcement. Meta says the model uses the Muse Model License with limited commercial terms; users should review the license for specific restrictions before commercial use.
Are the model’s performance and safety claims independently verified?
The supplied announcement reports Meta’s own evaluations and claims, but does not include benchmark scores or a complete safety report. Meta says a full evaluation report will be available soon; independent results are not provided in the source material.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
