Meta Superintelligence Labs Releases Its First AI Model, Muse Spark, for WhatsApp and Instagram

Meta Superintelligence Labs has launched Muse Spark, the first model in a new series created after a large-scale reorganization of the company's AI work. It is the lab's first major product result since Zuckerberg poured billions into a complete rethink of Meta's AI strategy, poached researchers from OpenAI, Google DeepMind and other leading AI labs, and built a structure aimed at competing head-on with the market leaders. Muse Spark already runs in the Meta AI app and on meta.

ai in the US. Over the next few weeks, the company plans to deploy the model across WhatsApp, Instagram, Facebook, Messenger and Meta Ray-Ban smart glasses. A rollout to other countries starts in parallel.

Specific markets have not been named, but given the scale of Meta's global footprint, it most likely means Europe, Latin America and Southeast Asia.

Muse Spark's positioning deliberately echoes Google's approach to Gemini: the model is described as built specifically for Meta's products. In theory, that means the architecture, context memory and system integrations are optimized for concrete scenarios inside the ecosystem, from messenger conversations to voice commands through smart glasses. The model is also available in a closed mode to some Meta partners, although the terms of that cooperation have not been disclosed.

Muse Spark is the first in a new series of models from Meta Superintelligence Labs, which in itself signals long-term plans. The company intends to build its own model stack rather than rely only on Llama and outside providers. Until now, Meta positioned Llama primarily as an open foundation for third-party developers; the open-weights strategy was meant to build an ecosystem and put pressure on competitors. But a product layer embedded in services with audiences in the billions calls for fundamentally different solutions: strict safety requirements, low latency, managed context and compliance with the platform's internal policy.

The launch comes amid fierce competition in integrated AI assistants. OpenAI is building GPT-4o into voice and search, Google is pushing Gemini 2.5 Pro as a universal model from Android to Workspace, and Apple is rolling out Apple Intelligence across its device lineup. With two billion daily active users, Meta has a unique advantage: built-in distribution through WhatsApp and Instagram, with no need to persuade people to install a separate app.

The key question is not whether Muse Spark will show up in the Instagram feed or in the glasses' voice assistant, but how noticeable the changes will be in everyday use. Photo caption generation, replies in direct messages, summaries of long chats, a personal assistant in AR glasses: each scenario places fundamentally different demands on latency, memory and response style. That Meta chose to build a specialized model instead of an adapted Llama fork points to serious intent.

For most users, the change is not yet visible: Meta AI in the apps looks the same as before. But under the hood, there has been a fundamental shift, from a strategy of taking open weights and adapting them to one of building a model from scratch for a specific product ecosystem. That is the line separating companies that embed someone else's AI in their products from those that build the full stack themselves. Meta has bet on the second path.

Meta has not published Muse Spark results on external benchmarks and has not disclosed technical details of the architecture. Practice will show the model's real level: how noticeably the AI features in Instagram and WhatsApp improve, and which partners decide to build products on it. The next few months of rollout will be the main indicator of whether the multibillion-dollar bet on its own AI stack was justified.

*Meta is designated an extremist organization and banned in Russia.