The eagerly awaited model, dubbed Muse Spark, was developed by Meta Superintelligence Labs, a newly formed team of top-tier AI researchers led by Chief AI Officer Alexandr Wang. Muse Spark will serve as the foundation for the Meta AI chatbot and, diverging from the company’s previous open-source strategy, is a closed model, meaning its architecture and code will not be publicly disclosed.
Following the announcement, Meta shares surged by 6% in New York.
Also Read: US court declines to block Pentagon’s Anthropic blacklisting for now
This model represents the initial significant test for Zuckerberg’s new AI lab, known as MSL. The Facebook founder brought Wang on board during a $14 billion investment into Scale AI last year after experiencing a series of setbacks that left him dissatisfied with the company’s competitive stance against rivals like OpenAI, Anthropic PBC, and Google. Since then, Zuckerberg has invested billions in AI talent and pledged tens of billions more towards infrastructure improvements, such as data centers, to close the gap.
Meta has sought to maintain agility in its AI division by granting researchers autonomy and reducing its traditionally heavy management structure, according to a company executive. Wang reportedly oversees around 100 direct reports, as mentioned by the unnamed executive, who shared insights on internal matters.
The executive recognized that Muse Spark may not perform as well as OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini in certain areas but emphasized that it is still early in the company’s development process. Meta characterized the model as “an early data point on our trajectory,” with several larger models in the works, as stated in a blog post. The Meta AI chatbot will remain free for users, but the company is contemplating the introduction of subscription fees down the line, disclosed the executive. The name Muse Spark alludes to the upcoming array of AI products: the model is designed to “spark,” or ignite, MSL’s initiatives, as described by the executive, who characterized it as compact and efficient. It is the first in a planned series of “Muse” models — a nod to Zuckerberg’s fascination with Greek and Roman mythology. Muse Spark was trained using various third-party open-source models, including Qwen from the Chinese giant Alibaba Group Holding Ltd., along with models from OpenAI and Google. The inclusion of a Chinese model contradicts the stance taken by US policymakers and tech leaders, who have expressed concerns about China’s standing in the rapidly evolving AI domain — citing national security issues. Numerous leading American AI companies have sought to restrict Chinese models similarly trained on US technology. “In line with industry practices, Meta employs techniques like distillation with rigorous safeguards to learn from openly available AI models and enhance our own,” explained a Meta spokesperson. Similar to its competitors, Muse Spark incorporates several reasoning modes: “Instant” mode, “Thinking” mode, and “Contemplating” mode — the latter delivering research-grade responses. The Meta executive indicated that the model excels in answering queries related to science, health, and math, but is less proficient in coding. The model will be utilized to enhance Meta’s suite of applications, spanning Instagram, Facebook, and WhatsApp. It already serves as a foundation for an early-stage shopping agent that aims to assist users in finding products like clothing or furniture more effectively. The agent is not yet capable of completing purchases on behalf of users and remains in the developmental phase, as clarified by the executive.
