Meta has released the weights for Muse Spark Open, a new open-weight model aimed primarily at software development and agentic workflows. The announcement is not just about the model: the company worked with Apple to enable local execution on a setup featuring two Mac Studio systems with M5 Ultra, using the latest version of MLX, Apple's open-source machine learning framework for Apple silicon.
It is a proposition aimed at a niche audience, at least in terms of the required hardware, but noteworthy for the message it conveys. Large frontier models are typically consumed via cloud infrastructure and data center GPUs; Meta and Apple, by contrast, point toward a path where a portion of those capabilities can be brought directly into a developer's workspace, without constantly sending data and prompts to a remote service.
Meta calls Muse Spark Open its most capable open-weight model developed in the United States. The term open-weight should be read precisely: the company makes the model parameters available, though not necessarily every element of the training process. The weights are distributed under the Muse Model License, which includes limited commercial terms while allowing many developers to use it free of charge. It is therefore not completely unrestricted software in an absolute sense, but a release that permits local deployment, inference, and customization by third parties.
A massive model with a smaller active footprint
Its sheer scale explains why running it requires an out-of-the-ordinary machine. Muse Spark Open features 816 billion total parameters, but activates 42 billion during processing—an architecture designed to rein in the effective computational load compared to a dense model of equivalent total size. Meta also reports a one-million-token context window and native multimodal capabilities, allowing it to process multiple content types within the same architecture.
In the presentation materials, the company positions the model for multi-step, continuous tasks: coding, working on complex software projects, multi-stage reasoning, and long-duration agentic activities. In this context, an agent does not simply produce a text response: it can plan a sequence of actions, use tools, and verify intermediate results. It is an area where control over the execution environment and processed data becomes just as critical as the quality of the generated response.
Meta reports having used Muse Code, powered by Muse Spark Open, to work on Ice Cubes, an open-source client for Mastodon: cited tests include bug fixes, new features, and the full compilation of the app. Another demo involves DaVinci Resolve, featuring frame-by-frame analysis of footage to identify the best moments and crop them into vertical format. These are company-provided examples rather than independent benchmarks, and should be viewed as indications of the workflows Meta is targeting rather than guarantees of the model's behavior in every scenario.
Why two Mac Studio systems come into play
Optimization for Apple silicon is the most concrete aspect of the announcement. Meta states that, via MLX, Muse Spark Open can run on two Mac Studio M5 Ultra systems. The point is not that a single desktop can host the entire model: according to the company, two systems are needed to hold the weights in available memory.
The setup leverages the unified memory of Apple chips. Meta notes that the integrated GPU of each M5 Ultra can access up to 512 GB of unified memory; two Mac Studio systems therefore reach a combined one terabyte. The company also cites 1.2 TB/s of memory bandwidth per chip and describes the two systems as a single compute resource for this workload. In other words, the goal is not to turn a standard Mac Studio into a workstation suited for any massive model, but to demonstrate that a model of this scale can run on-device on a specific, very expensive desktop architecture.
This distinction is essential. Running locally promises greater control over proprietary code, internal documents, and other sensitive data, while also reducing reliance on connectivity once the model is installed. But two Mac Studios in the required configuration represent an investment far beyond the average personal computer. The project is closer to development studios, researchers, and small teams looking for a private inference platform, rather than a consumer product ready for the general public.
MLX is central precisely because it is designed for the characteristics of Apple silicon, particularly the shared memory between CPU and GPU. Apple, for its part, has highlighted how the framework allows researchers and developers to experiment with models and methods locally in a private environment. For Meta, the advantage is pairing the release of the weights with an execution path that is already recognizable to the Mac ecosystem, rather than limiting the launch to traditional discrete NVIDIA GPU-based configurations.
An open release, but with multiple access paths
Those without the required hardware do not necessarily have to forego Muse Spark Open. The weights are available via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. The model is also accessible through hosted inference platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
Meta reports compatibility with SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp, and Unsloth, as well as the ability to perform further fine-tuning via PyTorch and TorchTitan. The stated support covers AMD, Apple, Intel, and NVIDIA silicon. For NVIDIA and AMD GPUs, the company cites self-optimizing Helion kernels integrated into SGLang and kernel-level performance improvements; these are figures released by Meta and not comparisons produced by independent labs.
The variety of channels is consistent with the open-weight approach: making the model deployable across different environments, from the cloud to a machine managed directly by the user. However, API access and local execution address different needs. The former offers operational simplicity at the cost of relying on an external provider and its associated usage fees; the latter provides control and privacy, but requires handling hardware, configuration, updates, and infrastructure responsibilities.
Safety, licensing, and the Apple-Meta relationship
Meta claims to have subjected Muse Spark Open to additional safety training precisely because of the distribution of its weights. The evaluation, the company explains, was conducted under the Meta Superintelligence Scaling Framework, and the model reportedly met the thresholds required for release. Detailed results are outlined in the model card, while a comprehensive evaluation report is expected later.
The availability of the weights inevitably broadens opportunities for community analysis and adaptation, but it also limits the direct control the developer retains post-download. It is one of the classic trade-offs of open-weight models: greater autonomy for developers and researchers, alongside an increased need to carefully evaluate licenses, application protections, and end uses.
Finally, the technical collaboration with Apple warrants attention due to the strategic weight of both groups. Meta brings a massive model and its agentic tools; Apple provides an optimized pathway on proprietary hardware and a framework designed for local experimentation. For now, the initiative remains a demonstration tailored to a specific configuration, rather than the announcement of a built-in macOS feature or an out-of-the-box assistant. Yet it shows that the Mac can become a viable destination even for local AI workloads that, until recently, would have been considered the exclusive domain of data centers.
