Meta Launches Closed-Source AI Model Muse Spark — First Product From New MSL Division

First Model After Major AI Reorganization

Meta has unveiled Muse Spark, the first language model produced by its newly established Meta Superintelligence Labs (MSL) research group. The model already powers the Meta AI app and website, with deployment to WhatsApp, Instagram, Facebook, Messenger, and the company's AI glasses planned for the coming weeks.

«Today we're introducing Muse Spark, our most powerful model yet, giving you a faster and smarter Meta AI. Muse Spark currently powers the Meta AI app and website and will be rolling out to @whatsapp, @Instagram, @facebook, @messenger, and AI glasses in the coming weeks.…» — Meta Newsroom (@MetaNewsroom), original post

Muse Spark represents Meta's first major LLM release since CEO Mark Zuckerberg launched a multi-billion-dollar restructuring of the company's AI division, aimed at closing the gap with leading competitors in artificial intelligence.

Why This Matters

The launch signals a fundamental shift in Meta's AI strategy. After years of championing open-source models through its Llama family, the company has released a fully closed-source product for the first time. This positions Meta to compete directly with ChatGPT, Claude, and Gemini in the proprietary AI segment.

The new MSL division is led by Alexandr Wang, whom Zuckerberg recruited as part of investment agreements with Scale AI totaling $14 billion. Following the team's formation, Meta has poured billions into talent acquisition and allocated tens of billions toward infrastructure development.

A Closed Model With Acknowledged Limitations

Muse Spark will serve as the foundation for the Meta AI chatbot. Unlike the company's previous open-source offerings, the new model is entirely proprietary.

Meta itself acknowledges that Muse Spark falls short of ChatGPT, Claude, and Gemini in certain areas. The company characterized the release as merely the "first step on the journey," with more powerful solutions currently in development.

The emphasis with Muse Spark is on efficiency and "competitive performance" across a range of tasks. Development took nine months, and leadership views it as a component of the company's refreshed AI strategy.

Training involved several third-party open-source tools, including Qwen from Chinese tech giant Alibaba, as well as solutions from OpenAI and Google. A company spokesperson said Meta employs distillation with strict safety measures, consistent with industry practices.

Three Thinking Modes and Monetization Plans

Muse Spark offers users three processing tiers:

• Instant — immediate responses;
• Thinking — a reasoning mode;
• Contemplating — deep reasoning that delivers research-grade answers.

The Contemplating mode employs a group of agents for "parallel thinking." Its integration into the Meta AI app and web experience will be rolled out gradually. Users can switch between modes depending on the complexity of their query.

Meta is also testing new monetization channels. Third-party developers will be offered access to Muse Spark's underlying technology via API. Additionally, an updated version of Meta AI will include a dedicated "Shopping" mode designed to help users browse and purchase clothing.

Earlier in February, Meta introduced an AI feature called Dear Algo for Threads, enabling users to personalize content recommendation algorithms.