Meta AI Unveils Muse Spark Open: A New Era for Open-Weight Models

October 15, 2026 | 4 minute read

Meta AI Research has officially launched Muse Spark Open, currently the most sophisticated open-weight model developed in the US. This release marks a milestone in accessibility, bringing high-tier agentic capabilities directly to the desktop environment—specifically for users with Apple M5 Ultra Mac Studios.

🚀 Model Architecture & Capabilities

Muse Spark Open is designed as a natively multimodal powerhouse. It is engineered specifically to handle complex software engineering, multi-step logical reasoning, and "long-horizon" tasks.

Technical Specifications

Feature	Specification
Total Parameters	$816\text{B}$
Active Parameters	$42\text{B}$
Context Window	$1\text{M}$ tokens
Primary Focus	Coding & Agentic Workflows

"We believe the future is for everyone, and that includes the freedom to run advanced models wherever, including at home."

💻 Optimized for Apple Silicon

For the first time, a model of this magnitude can be operated on a desktop. By utilizing the latest MLX framework and two Apple M5 Ultra Mac Studios, the hardware acts as a localized supercomputer.

The Hardware Synergy

The efficiency stems from Apple's unified memory architecture. With each M5 Ultra supporting up to $512\text{ GB}$ of integrated GPU memory, a dual-machine setup provides a massive $1\text{ TB}$ of unified memory. Furthermore, the memory bandwidth of $1.2\text{ TB/s}$ ensures that the model can read the necessary token slices with minimal latency.

Real-World Agentic Testing

Meta tested the model using Muse Code to perform end-to-end professional tasks:

iOS Development: Successfully implemented bug fixes and new features for Ice Cubes (an open-source Mastodon client), including a full build.
Video Production: Automated DaVinci Resolve by analyzing footage frame-by-frame to identify and crop the best moments for vertical formats.

🛠️ Deployment & Integration

Muse Spark Open is designed for maximum flexibility across various silicon architectures, including NVIDIA, AMD, Intel, and Apple.

Where to Access the Model

API Platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
Weight Downloads: Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.

Supported Inference Engines

The model is compatible with a wide array of engines: SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth.

Performance Boosts

The open-weight release includes specialized, auto-tunable Helion GPU kernels within SGLang.

# Example of potential performance gain
kernel_speedup = "1.03x to 3.95x" # On NVIDIA and AMD GPUs
end_to_end_gain = "3%" # Specifically for AMD MI350

🛡️ Safety and Licensing

Meta is committed to the idea that ~~closed models are the only way to be safe~~ open weights actually help advance safety research across the entire industry.

Safety Measures:

Post-Training: The model underwent additional safety tuning specifically for open-weight distribution.
Framework: Evaluated via the Meta Superintelligence Scaling Framework to mitigate severe risks.
Transparency: Detailed findings are available in the Muse Spark Open model card.

Licensing: The weights are released under the Muse Model License. While it contains limited commercial terms, it remains free for a vast majority of developers.

🏁 Getting Started

Ready to build? Here is how to dive in:

Customization: Use PyTorch’s TorchTitan to fine-tune the model for your specific needs.
Apple Silicon Setup: Use MLX to pull weights directly from Hugging Face.
Community: Join the conversation on the Hugging Face Forums.

Quick Links:

Weights: Hugging Face Meta Collection
Tooling: MLX Adapter for Mac Studio.