Apple et Meta s'associent pour apporter Muse à code source ouvert sur les Mac

Meta has launched Muse Spark Open, an open-weight AI model described as the most capable developed in the United States. The model, which includes 816 billion parameters and supports a 1 million token context window, is optimized for local execution on Apple’s Mac computers. According to Meta, the model demonstrates superior performance compared to other prominent open-weight models, based on evaluations conducted by the company. Designed to function natively on Macs, Muse Spark Open is intended for complex coding tasks and agentic operations, where AI agents perform multiple steps or tasks autonomously. The model can be executed locally on Apple’s Mac Studio computers equipped with the M5 Ultra chip. However, due to its size, it demands substantial computational resources, requiring at least 1 terabyte of RAM. For optimal performance, users would need two or more Mac Studios linked via Apple’s MLX framework, which distributes processing across multiple machines. Meta has partnered with Apple to enable the model to operate on its hardware, marking a notable shift in relations between the two companies. Previously, Apple and Meta had been competitors in areas such as privacy policies and device ecosystems. Now, they collaborate to bring powerful AI capabilities directly to personal computers, offering an alternative to traditional GPU-based solutions and reducing reliance on proprietary models. According to Meta, Muse Spark Open has already been applied in practical scenarios, including fixing bugs and implementing new features in a Mastodon client, as well as assisting with video editing software like DaVinci Resolve. The model is being made freely available through various API platforms, with a license permitting commercial use. The release of Muse Spark Open follows recent developments in the global AI landscape, where several major firms have introduced powerful open-weight models. These include updates from Nvidia’s Nemotron series and new releases from U.S.-based Reflection AI and French firm Mistral. Despite these advancements, Muse Spark Open appears to outperform many of its counterparts in terms of parameter count and overall capability. Meta’s AI chief, Alexandr Wang, highlighted the significance of the collaboration with Apple, noting that the partnership had been ongoing for several months. A former Apple executive, Daniel Gross, has played a pivotal role in facilitating the relationship between the two companies. In addition to its technical specifications, Muse Spark Open represents a broader trend in the AI industry, where companies are increasingly seeking partnerships to enhance their offerings and reduce dependency on specific technologies or vendors. While Apple and Meta continue to compete in certain areas, their joint effort underscores the growing importance of AI integration in personal computing environments. The model’s availability is set to expand further, with plans to deploy it through cloud services and developer tools. This move allows businesses and individuals to leverage the model’s capabilities without relying solely on cloud infrastructure, thereby enhancing flexibility and reducing costs associated with AI deployment. Meta’s decision to return to open-weight models reflects a strategic pivot following earlier focus on closed models. With Muse Spark Open, the company aims to provide greater accessibility and customization options for developers and organizations looking to integrate advanced AI functionalities into their workflows.