Meta Launches Open AI Model to Challenge Chinese Rivals

Meta has unveiled Muse Spark Open, a new open-weight artificial intelligence model that the company says can rival, and in some tests outperform, leading Chinese alternatives, as it steps up efforts to strengthen the United States’ position in a rapidly evolving AI market. Released on Thursday, October 15, the model is designed for complex coding, multi-step reasoning and AI agents capable of carrying out tasks with limited human intervention.

Meta describes Muse Spark Open as the most capable open-weight model developed in the United States to date. It has 816 billion parameters in total, with 42 billion active during operation, and supports a context window of up to one million tokens. These features are intended to help it handle lengthy instructions, complex software projects and tasks requiring multiple stages of planning and execution. The performance claims, however, are based partly on evaluations published by Meta, and independent testing will be important in determining how it compares with competing systems in everyday use.

Unlike fully closed AI systems, an open-weight model makes its trained parameters available for developers to download, run and customize. Meta has made Muse Spark Open’s weights available through platforms including Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, while several providers are offering access through their APIs. The company is also releasing the model under its own licence, which permits many uses but includes limits on commercial terms. It is therefore more precise to describe the model as open-weight rather than assume that all its training data, development processes and underlying code are fully open-source.

The model’s capabilities come with substantial hardware demands. Meta says it has optimized Muse Spark Open to run locally using two Mac Studio computers equipped with Apple’s M5 Ultra chips, which together can provide enough unified memory to accommodate the model. The collaboration is notable because Meta and Apple have often competed over consumer technology and digital platforms. Running a model locally can give developers greater control over their systems and data, but the equipment required means that the most powerful configuration will remain beyond the reach of many individual users.

The launch comes as Chinese developers have gained ground in the open-weight AI market with models such as Alibaba’s Qwen, Moonshot AI’s Kimi and Z.ai’s GLM family. These systems have attracted developers seeking alternatives to proprietary services, particularly where customization, cost and control over deployment matter. Meta’s new release is intended to give US developers a stronger option in that segment, although the company’s claims of competitive performance will need to be assessed against independent benchmarks and practical results.

Meta’s return to open-weight development also follows a broader overhaul of its AI strategy. In August, chief executive Mark Zuckerberg called for changes to US policies governing open AI development, arguing that American companies faced disadvantages compared with competitors in China. The company has since pursued both downloadable models and more tightly controlled systems, reflecting an effort to compete across different parts of the AI market rather than relying on a single approach.

Earlier this week, Meta introduced Muse Wonder and Muse Wonder Full, two more advanced models aimed at personal and professional AI agents. Those releases were positioned against products from OpenAI and Anthropic, while Muse Spark Open targets developers who want to download and adapt model weights for their own applications. Together, the launches illustrate Meta’s attempt to build a broader AI portfolio spanning consumer tools, business services and models that outside developers can customize.

The stakes extend beyond the rivalry between US and Chinese technology companies. Open-weight models can widen access to advanced AI by allowing businesses, researchers and independent developers to adapt systems to local needs instead of depending entirely on a single provider’s platform. At the same time, releasing powerful models for broader use raises questions about security, safeguards and the resources required to run them responsibly. Muse Spark Open’s uptake will therefore depend not only on its benchmark results, but also on whether developers find it practical, affordable and reliable for real-world work.