In 2025, Mark Zuckerberg invested billions in establishing the Meta Superintelligence Labs (MSL). Meta appointed Alexandr Wang to lead MSL following the acquisition of Scale AI for over $14 billion. Recently, MSL introduced Muse Spark, its latest AI model, which Meta claims to be their most advanced creation yet.
According to Alexandr Wang, Muse Spark was developed from scratch within a span of nine months. This new model will now drive Meta AI across multiple platforms like Instagram, WhatsApp, and Facebook.
There are indications that Muse Spark may have been the model previously known as Avocado during testing by MSL.
Muse Spark belongs to Meta’s new Muse family of AI models, intended to replace the older Llama models. While Spark is not Meta’s top-tier model, it is described as a small, fast, and capable model designed for complex problem-solving in science, math, and health. Meta asserts that Muse Spark can rival models like Anthropic’s Claude Opus 4.6 and OpenAI’s GPT 5.4 in various benchmarks, including multimodal reasoning, health applications, and agentic AI.
Alexandr Wang highlighted that Muse Spark’s contemplating mode can handle multiple agents simultaneously for complex tasks. He mentioned that during testing, the model demonstrated competitiveness with other advanced reasoning models such as Gemini Deep Think and GPT Pro.
Muse Spark’s functionalities include understanding the surrounding environment, solving intricate scientific and mathematical problems, and retrieving information from conversations within Meta’s apps. The AI is tailored for the vast user base of 3 billion people engaging with Meta apps daily.
Muse Spark powers an updated Meta AI version accessible via the Meta AI app and website, with plans for expansion to Facebook, Instagram, WhatsApp, Messenger, and Meta’s smart glasses. Additionally, a new shopping mode offers personalized recommendations based on creators and brands.
In contrast to the open-sourced Llama models, Muse Spark is currently closed source, although Meta hints at potential future open-sourcing. The model is presently available as a private API preview for select partners. Alexandr Wang indicated that this is just the initial phase, with larger models under development and scaling infrastructure in progress.
Alexandr Wang claimed that Muse Spark exhibited robust refusal behavior in high-risk domains like biological and chemical weapons during extensive safety assessments compared to similar models from Anthropic and OpenAI.