НВ (Новое Время)

Meta Launches New AI Model Muse Spark, Requiring Ten Times Fewer Resources

Meta, a company renowned for its technological innovations, has announced the launch of a groundbreaking artificial intelligence model called Muse Spark, which reportedly consumes over ten times fewer computational resources compared to its predecessors.

Meta, a leader in technological advancements, has unveiled its latest artificial intelligence model, Muse Spark. According to company representatives, this model is revolutionary as it integrates text, images, and tools into a single system while requiring more than ten times fewer computational resources compared to previous models.

Interesting Engineering reports that Muse Spark was introduced as the first development from Meta's division known as Superintelligence Labs. This model employs multimodal thinking, enabling it to perform tasks through parallel working agents. This means that Muse Spark can simultaneously process textual information, images, and various tools, significantly enhancing its efficiency.

At Meta, officials emphasize that the new model assists in tackling complex tasks more effectively, although they acknowledge that the system still has its shortcomings. Muse Spark reflects the overall trend in artificial intelligence development, where modern models not only generate text but also analyze visual and real-world data.

One of the key features of Muse Spark is its ability to analyze images, solve problems in STEM disciplines, and recognize objects in context. The model also supports step-by-step explanations based on images, which Meta refers to as a "visual chain of thought." This allows users to receive practical instructions, such as for repairing equipment, complete with visual prompts.

Additionally, Muse Spark is capable of creating interactive content, including simple games upon user request. However, despite these capabilities, such functions in the AI sector still operate unevenly. Despite progress, the stability of performance in real-world conditions remains lower than the results achieved in testing.

One of Muse Spark's significant innovations is the "thinking mode" feature, which activates multiple thinking agents simultaneously for more effective problem-solving of complex tasks. Meta reports that Muse Spark scored 58% on the Humanity’s Last Exam and 38% on tasks from FrontierScience Research. These metrics assess the model's ability for complex thinking, but they are challenging to compare across different systems due to varying evaluation methods.

The company also notes that the model has become more reliable while maintaining a diversity of responses. They claim that Muse Spark performs better on tasks not included in the training data, although independent verification of these claims is still lacking.

The development of Muse Spark was accompanied by significant changes in the company’s infrastructure. According to Meta, over the past nine months, the company has restructured its model training process, focusing on architecture, optimization, and data quality. This, they assert, has allowed them to achieve similar results with over ten times fewer computational costs compared to the previous model, Llama 4 Maverick.

Meta also emphasizes that reinforcement learning remains a key element of their approach. With the scaling of training, the company observes consistent improvements in both training and testing of models. Muse Spark is viewed as a step toward creating what is termed "personal superintelligence"—systems capable of understanding the user's environment and providing individualized assistance.

One of the initial applications for Muse Spark is anticipated to be in medicine. For this purpose, training data was developed in collaboration with physicians to enhance explanations of medical topics. Currently, Muse Spark is available on Meta's platforms and has limited access through an API for developers.

The launch of this new model signifies shifts in competition within the artificial intelligence sector. Companies are focusing not only on creating smarter models but also on developing systems capable of functioning in real-world scenarios, although issues of reliability and validation remain open questions.