Meta's Superintelligence Lab Ships First Internal AI Models as Zuckerberg Eyes Frontier Competition

Meta's CTO Andrew Bosworth confirmed in January 2026 that the company's Superintelligence Labs has delivered its first internal AI models. The lab, led by former Scale AI CEO Alexandr Wang and approximately six months old at the time of the announcement, produced two models: Avocado, a large language model focused on coding and reasoning, and Mango, a multimodal model for image and video generation.

The Models

Avocado targets the coding and reasoning capabilities that have become the primary competitive battleground among frontier AI labs. Coding performance is measurable, reproducible, and directly tied to commercial value — developers who can generate, debug, and refactor code with AI assistance represent one of the largest paying customer segments for AI companies. Reasoning — the ability to solve multi-step problems, follow complex instructions, and maintain logical consistency — is the capability that separates models that are useful for simple queries from models that can handle enterprise workflows.

Mango covers the multimodal space, generating images and video from text prompts. Both models reportedly outperform earlier Llama models on internal benchmarks, though Meta has not published external benchmark results or disclosed the models' parameter counts, training data composition, or architectural details.

Public release of both models is targeted for the first half of 2026. Whether they will be released as open weights — Meta's established approach with the Llama series — or kept proprietary has not been confirmed.

Strategic Shift

The Superintelligence Labs represents a deliberate departure from Meta's previous approach to AI development. The Llama series was developed within Meta's FAIR (Fundamental AI Research) lab and released as open-weight models, establishing Meta as the leading corporate contributor to open-source AI. The decision to create a separate lab, recruit a high-profile external leader in Alexandr Wang, and frame the unit's mission around "superintelligence" signals that Meta is pursuing frontier capabilities on a separate track from its open-source efforts.

Meta has described 2026 and 2027 as "make-or-break years" for AI in everyday life — an internal framing that reflects the broader industry consensus that the current period will determine which AI products become deeply embedded in consumer and enterprise workflows. Zuckerberg has made AI central to Meta's public strategy, frequently referencing the company's massive infrastructure investments and its vision for AI assistants integrated across WhatsApp, Instagram, Facebook, and the standalone Meta AI product.

Implications for Open Source

If Meta releases Avocado and Mango as open weights, the impact on the broader AI ecosystem could be substantial. Each Llama release has rapidly become the foundation for hundreds of fine-tuned variants developed by independent researchers and companies. A frontier-class open-weight model from Meta would compress the capability gap between proprietary API-only models and freely available alternatives, affecting the economics of AI deployment for organizations that currently pay API fees for top-tier access.

Whether Meta will maintain its open-weights approach for the Superintelligence Labs output remains an open question. The competitive dynamics of the current AI market create tension between openness and keeping frontier capabilities proprietary long enough to extract commercial value.