Meta announced the release of Muse Spark Open, an open-weight artificial intelligence model that the company states is the most capable of its kind from a United States laboratory. The model, developed by Meta Superintelligence Labs, is designed for complex coding and reasoning tasks and possesses multimodal capabilities, processing text, images, audio, and video. Meta claims Muse Spark Open rivals, and in some instances surpasses, similar models developed by leading Chinese companies.
Muse Spark Open features an 816 billion parameter architecture, employing a mixture-of-experts (MoE) approach where approximately 42 billion parameters are active per task. This design allows for efficient processing, with the model capable of handling up to one million tokens in its context window. Meta has emphasized the model's proficiency in agentic coding workflows, including reading codebases, writing and debugging code, and utilizing tools. The model also integrates additional safety post-training to enhance its robustness against misuse.
In a notable development, Meta has partnered with Apple to optimize Muse Spark Open for Apple's hardware. The model can reportedly run on a desktop setup utilizing two Mac Studios equipped with Apple's M5 Ultra chips, leveraging a combined 1TB of unified memory. This collaboration allows for high-performance AI processing directly on desktop computers without relying on large server infrastructure. Meta CEO Mark Zuckerberg acknowledged the partnership, thanking Apple CEO John Ternus for their work on the project.
The release positions Meta to compete more directly in the open-weight AI model market, an area where Chinese developers such as Alibaba, Moonshot AI, and Z.ai have established a significant presence. Meta's previous AI efforts included the release of Muse Glimmer in August, a smaller dense model, and a closed model announced shortly before Muse Spark Open that reportedly rivaled advanced versions of ChatGPT and Claude. The company has expressed a commitment to open source and open science in advancing artificial intelligence.
Muse Spark Open is available on platforms including Hugging Face, and supports various inference engines such as SGLang and vLLM, with optimizations for NVIDIA, AMD, and Intel hardware. The model's multimodal perception allows it to process visual and auditory inputs within a real execution environment, moving beyond scripted steps for tasks like building from screenshots or analyzing video clips. Meta has also introduced draft models for faster token generation, designed to verify multiple tokens in a single pass.