Meta has launched a new artificial intelligence model designed to prioritize speed and efficiency over sheer size, marking a strategic shift as the company attempts to reclaim ground in a landscape currently dominated by OpenAI, Google, and Anthropic. The new model, called Muse Spark, represents the first major output from Meta Superintelligence Labs, a specialized unit established to accelerate the company’s development cycle and move beyond the limitations of previous open-source efforts.
The move comes as Meta debuts new AI model capabilities aimed at integrating deeply into its existing ecosystem of apps and hardware. Originally developed under the codename “Avocado,” Muse Spark is not positioned as a “frontier” model in terms of scale, but rather as a high-performance tool capable of reasoning through complex scientific, mathematical, and health-related queries while requiring significantly less computing power than its predecessors.
This pivot follows a period of perceived stagnation for Meta’s AI ambitions. After the debut of several open-source models last April failed to gain significant traction among developers, CEO Mark Zuckerberg shifted the company’s focus toward a more agile, proprietary approach. Central to this transformation is Alexandr Wang, the founder and former CEO of Scale AI, who joined Meta in June following a $14.3 billion investment in Scale AI. Under Wang’s leadership, Meta Superintelligence Labs has spent nine months rebuilding the company’s AI stack from the ground up.
Meta CEO Mark Zuckerberg makes a keynote speech at the Meta Connect annual event at the company’s headquarters in Menlo Park, Calif., on Sept. 25, 2024.
Manuel Orbegozo | Reuters
A Strategic Shift Toward Efficiency
For years, the AI arms race was defined by the size of the model—the “bigger is better” philosophy. Yet, Meta is now betting on “competitive performance” through efficiency. In a technical blog post, the company revealed that Muse Spark can match the capabilities of older, midsize Llama 4 variants while utilizing an order of magnitude less compute. This allows the model to be deployed more widely across mobile devices and wearable tech without the massive latency or energy costs associated with larger models.
While Meta continues to invest in “long-horizon agentic systems” and coding workflows—areas where it acknowledges performance gaps—Muse Spark is designed to excel in multimodal perception and reasoning. This means the AI can better understand and process information across text, images, and audio simultaneously, providing a more seamless user experience.
The market reacted positively to the announcement, with Meta’s stock rising nearly 9% on Wednesday. While this rally coincided with a broader market jump following President Donald Trump’s announcement regarding the suspension of attacks on Iran, the surge underscores investor confidence in Meta’s renewed AI trajectory.
Integrating AI into the Meta Ecosystem
Unlike the Llama series, which championed an open-source philosophy, Muse Spark is currently proprietary. Meta has indicated it may open-source future versions, but for now, the focus is on tight integration and monetization. The model is already powering the digital assistant in the standalone Meta AI app and website, with a wider rollout planned for Facebook, Instagram, WhatsApp, and Messenger in the coming weeks.
The model will also be integrated into Ray-Ban Meta AI glasses, bringing advanced reasoning and multimodal capabilities to the company’s wearable hardware. Meta intends to use Muse Spark to power its “Vibes” AI video feature, reducing its current reliance on third-party models from providers like Black Forest Labs.
To enhance user interaction, Meta is introducing specialized modes within the AI app that allow users to toggle based on the complexity of their needs:
| Mode | Primary Use Case | Key Capability |
|---|---|---|
| Quick | Simple questions | Rapid, low-latency responses |
| Complex | Legal/Nutritional analysis | Deep document and image reasoning |
| Contemplating | Extreme reasoning tasks | Parallel processing via AI agents |
| Shopping | Commerce and styling | Creator-led brand storytelling |
The “Contemplating” mode is particularly notable for its use of a “squad of AI agents” that reason in parallel. This architectural choice is a direct attempt to compete with the high-conclude reasoning capabilities found in Google’s Gemini Deep Think and OpenAI’s GPT Pro.
Alexandr Wang speaks on CNBC’s “Squawk Box” outside the World Economic Forum in Davos, Switzerland, January 23, 2025.
Gerry Miller | CNBC
The High Cost of Competition
The financial stakes of this transition are staggering. Meta is aggressively ramping up its infrastructure spending to keep pace with other “hyperscalers.” According to the company’s latest earnings report, AI-related capital expenditures for 2026 are projected to fall between $115 billion and $135 billion—nearly double the spending of the previous year.
This spending spree is driven by a massive projected growth in the generative AI sector. Research from Grand View Research estimates the global market will grow by more than 40% annually, climbing from approximately $22 billion in 2025 to nearly $325 billion by 2033.
To offset these costs, Meta is exploring new revenue streams. The company has launched a private API preview of Muse Spark for select partners, with plans to eventually offer paid API access to a broader audience of third-party developers. This marks a significant departure from the company’s previous strategy of giving away model weights to foster a developer ecosystem, shifting instead toward a “model-as-a-service” business model.
Disclaimer: This article contains information regarding financial projections and stock market movements; We see intended for informational purposes and does not constitute investment advice.
The next phase for Meta will be the gradual rollout of the “Contemplating” and “Shopping” modes to the general public, alongside the integration of Muse Spark into its suite of social messaging apps. The company is expected to provide further updates on the performance of these agentic systems in its next quarterly earnings call.
We want to hear from you. Do you think a shift toward proprietary, efficient models is the right move for Meta, or should they have stuck with an open-source path? Share your thoughts in the comments below.
