Meta Platforms Inc. launched its latest artificial intelligence model, Muse Spark, on April 15, 2026, marking the debut of the highly anticipated Avocado series. Developed by the company’s newly formed superintelligence division, the model is designed to replace existing Llama-based systems across WhatsApp, Instagram, and Facebook. This release represents a significant pivot in Meta’s strategy as it shifts from open-source research toward proprietary, high-performance frontier models.
The Genesis of Meta Superintelligence Labs and Avocado
The development of Muse Spark was led by Meta Superintelligence Labs (MSL), a specialized unit established in 2025 to accelerate the company’s quest for advanced cognitive computing. The division is headed by Alexandr Wang, the CEO of Scale AI, who took on the leadership role following a massive strategic partnership between the two companies. The Avocado project was kept under strict internal secrecy for over nine months as Meta overhauled its underlying data pipelines and compute architecture.
To support this new era of AI, Meta utilized its Prometheus supercluster, one of the world’s most powerful AI compute clusters. The company has significantly increased its capital expenditure, following a roadmap that involves spending between $115 billion and $135 billion on AI infrastructure by the end of 2026. This investment highlights the scale of the transition from general-purpose models to specialized “personal superintelligence” systems.
Technical Architecture and Multimodal Capabilities
Muse Spark is built from the ground up as a natively multimodal model, meaning it does not rely on separate plugins to process different types of data. It can simultaneously understand text, voice, and visual information, allowing for more natural interactions. This capability is particularly crucial for Meta’s Ray-Ban smart glasses, as the AI can now “see” the user’s environment in real-time and provide contextual assistance based on physical surroundings.
Advanced Reasoning and Contemplating Mode
A standout feature of the new model is its Contemplating Mode, a reasoning framework that allows the AI to “think” before responding. Instead of generating the most likely next word immediately, the system runs multiple sub-agents in parallel to evaluate various solutions for complex problems. This architecture enables Muse Spark to tackle high-stakes tasks in fields like scientific research, advanced mathematics, and healthcare diagnostics with much higher accuracy than previous iterations.
Strategic Shift Toward Proprietary Frontier Models
The launch of the Avocado series marks a departure from Meta’s historical commitment to the open-source community through its Llama (Large Language Model Architecture) family. While Llama, which debuted in 2023, remains available for developers and researchers, Muse Spark is a closed-source, proprietary model. This change allows Meta to exercise greater control over its core technology and protect its intellectual property in an increasingly competitive global market.
By moving to a closed-source model, Meta aims to create a deeper “platform lock-in” for its billions of users. The proprietary nature of the Avocado series also addresses growing security concerns regarding the misuse of open-weights models. This strategy aligns Meta more closely with competitors like OpenAI and Google, who primarily offer their most advanced models through controlled interfaces and paid APIs.
Integration Across the Meta Ecosystem
Meta has already begun the rollout of Muse Spark to its flagship Meta AI assistant. Over the coming weeks, the model will be integrated into the messaging and social feeds of WhatsApp, Instagram, and Messenger. This transition will see Muse Spark replace the older Llama-based chatbots, offering users a more personalized experience by leveraging their social graph, which represents the network of relationships and interactions they have across Meta’s platforms.
| Feature | Llama Series (Previous) | Muse Spark (New) |
|---|---|---|
| Availability | Open-weights / Open-source | Closed-source / Proprietary |
| Development | Meta AI Research | Meta Superintelligence Labs |
| Multimodality | Often required separate adapters | Natively multimodal |
| Reasoning | Standard sequential processing | Multi-agent “Contemplating Mode” |
| Focus | Foundational building blocks | Integrated personal superintelligence |
Vision for Personal Superintelligence
The launch of Muse Spark is the first step toward Mark Zuckerberg’s vision of “Personal Superintelligence.” Rather than focusing solely on artificial general intelligence (AGI), which aims to mirror human-level intelligence across all tasks, Meta’s goal is to create a deeply personalized companion. This “companion” is designed to understand a user’s habits, preferences, and social context to help them be more productive and creative.
Looking ahead, Meta plans to follow up with “Mango,” another internal codename for a model focused specifically on high-fidelity video and media generation. Together, these initiatives aim to transform Meta from a social media conglomerate into an AI-first company where every interaction, from private messaging to online commerce, is facilitated by an intelligent, multimodal agent.
Key Takeaways
- Meta launched Muse Spark, the first artificial intelligence model from its Avocado series, on April 15, 2026.
- The model was developed by Meta Superintelligence Labs (MSL), which is led by Alexandr Wang.
- Muse Spark features a new Contemplating Mode that uses parallel sub-agents to solve complex reasoning problems in math and science.
- Meta is transitioning from the open-source Llama model family to a closed-source, proprietary strategy for its frontier AI.
- The new model will power the AI assistants on WhatsApp, Instagram, Facebook, and Ray-Ban smart glasses.
- Meta’s AI roadmap includes a capital expenditure of up to $135 billion by the end of 2026 to support its Prometheus supercluster.
