Meta has released Muse Spark AI, the first model to emerge from its newly restructured superintelligence division, a unit that CEO Mark Zuckerberg spent billions of dollars building over the past year to close the gap with competitors.
The model is available now on Meta’s standalone AI app. It will roll out to WhatsApp, Instagram, and Meta’s AI smart glasses in the coming weeks.
Performance and Limitations of Muse Spark AI
Writing and reasoning benchmarks show Muse Spark AI performs significantly better than Meta’s previous models. Results place it close to top offerings from Google, OpenAI, and Anthropic, though it still trails those rivals in coding ability, an area that has become a central focus of the broader AI race.
A Closed Model, Departing From Meta’s Open-Source History
Unlike Meta’s earlier models, including the Llama series, Muse Spark AI uses private code and is not open-source. The company said it may release parts of the model as open-source in the future. This marks a notable shift in strategy, as Meta had long positioned open-source artificial intelligence as a competitive advantage and a point of differentiation from rivals.
The decision signals that Meta now treats AI leadership as too valuable to share freely while it works to close gaps in coding performance.
The Team Behind the Model
Muse Spark AI was built by a new superintelligence lab led by Alexandr Wang, 29, who was hired as Meta’s chief AI officer. The next model in development is known internally as Watermelon. Meta’s previous flagship model, Llama 4, disappointed upon release last year and triggered an executive restructuring that led to the formation of this new division.
“Meta forecasts spending up to $135 billion this year, nearly twice the $72 billion spent last year, with much of that directed toward AI infrastructure.”
Meta Financial Guidance, 2026
Massive Capital Commitments Back the Push
Zuckerberg has pledged a $600 billion investment in new data centers as part of Meta’s effort to win the AI race. The company’s projected 2026 spending of up to $135 billion represents a sharp increase from the $72 billion it spent in 2025. Most of that capital is directed at AI infrastructure and compute capacity.
The scale of investment reflects how seriously Meta is treating the competitive threat from OpenAI and Google, both of which have maintained leads in model capability and developer adoption. Meta’s apps, including WhatsApp and Instagram, give it a distribution advantage that few AI companies can match, and Muse Spark AI is the first model designed to fully exploit that reach.
What Comes Next
With Watermelon already in development, Meta appears to be moving toward a faster release cadence. The company’s ability to close the coding gap will likely determine whether Muse Spark AI is seen as a turning point or a transitional step. Analysts and developers will watch closely as the model reaches WhatsApp and Instagram users, platforms with billions of active accounts worldwide.
