Meta on Wednesday unveiled Muse Spark, the first AI model produced by its Superintelligence Labs division — and in a dramatic strategic pivot, the company is keeping it proprietary. The move marks Meta's clearest break yet from the open-source philosophy that defined its Llama model family and signals its intent to compete head-on with OpenAI, Anthropic, and Google in the commercial AI race.
Built From the Ground Up in Nine Months
Code-named Avocado during development, Muse Spark was built over nine months by a team led by Alexandr Wang, whom Meta brought in as its first-ever chief AI officer after acquiring a 49% nonvoting stake in Wang's company Scale AI for $14.3 billion in June 2025.
According to Meta's announcement, the Superintelligence Labs team "rebuilt our AI stack from the ground up, moving faster than any development cycle we have run before." The result is what Meta describes as a "small and fast" model that can reason through complex questions in science, math, and health while using roughly an order of magnitude less compute than its older midsize Llama 4 variants.
Capabilities and Deployment
Muse Spark accepts text, voice, and image inputs and features three operating modes: an Instant mode for quick responses, a Thinking mode for deeper individual reasoning, and a Contemplating mode that orchestrates multiple AI subagents in parallel for the most complex reasoning tasks. The model also supports visual coding — generating websites and mini-games — and includes a new shopping mode that integrates styling recommendations from creators.
The model now powers the Meta AI app and the meta.ai website, with rollouts to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban Meta AI glasses planned in the coming weeks. API access is currently limited to a private preview for select partners.
Competitive Performance, Closed Access
Meta says Muse Spark is competitive with frontier models from OpenAI, Anthropic, and Google across a range of benchmarks, though it acknowledged the model does not surpass them across the board. The efficiency claims are notable: if validated independently, achieving comparable performance at a fraction of the compute cost could reshape pricing dynamics in the API market.
The decision to keep Muse Spark proprietary is the story's most consequential detail. Meta had positioned itself as the champion of open-weight AI with Llama, winning developer loyalty and challenging the closed-model paradigm. The company says it hopes to open-source future versions of the Muse series, but for now, the model's design and weights remain under wraps.
Market and Industry Impact
Wall Street responded enthusiastically: Meta's stock surged roughly 8% following the announcement. The company also indicated it could spend between $115 billion and $135 billion on capital expenditure in 2026, predominantly on AI infrastructure and data centers.
For the broader industry, Muse Spark raises a pointed question: if even Meta — open-source AI's most powerful advocate — is going proprietary, what does that signal about the economics of frontier model development? The answer may depend on whether Muse Spark's efficiency gains hold up under independent scrutiny and whether Meta follows through on its open-source commitments for future models.
