Meta Platforms is making a bold strategic shift in its artificial intelligence efforts by developing a proprietary, closed-source AI model internally codenamed “Avocado,” which is slated for release sometime in the spring of 2026. This represents a significant departure from CEO Mark Zuckerberg’s longstanding advocacy for open-source AI technologies, a philosophy that had defined much of Meta’s approach in recent years through models like Llama. Instead of freely sharing the full capabilities of Avocado with the public and developers worldwide, Meta plans to maintain tight control over the model, potentially monetizing access through licensing deals, subscriptions, or enterprise partnerships—much like competitors Google with its Gemini family and OpenAI with GPT series have done successfully.
The decision underscores a pragmatic evolution in Meta’s thinking, driven by the realities of a fiercely competitive AI landscape where proprietary advantages can translate directly into revenue streams and market dominance. Bloomberg’s detailed reporting on Tuesday first brought this pivot to light, highlighting how Meta’s leadership sees closed models as essential for protecting intellectual property, accelerating internal advancements, and capturing a slice of the burgeoning AI services market projected to reach hundreds of billions in value over the next few years. This turnaround comes on the heels of the underwhelming performance of Meta’s Llama 4 model, which launched in April 2025 amid high expectations but quickly faced backlash from the developer community. Critics pointed to subpar results on standard benchmarks like MMLU for reasoning, HumanEval for coding, and GSM8K for math, where Llama 4 lagged behind frontrunners from rivals.
Adding fuel to the fire, developers alleged that Meta had showcased inflated metrics during public demos using an enhanced internal variant, while the downloadable open-source version delivered noticeably weaker outputs—a discrepancy that eroded trust and sparked widespread frustration online and in tech forums. In response, Zuckerberg took decisive action: he sidelined key team members involved in the Llama 4 project, restructured development workflows, and launched an aggressive talent acquisition drive. This included multimillion-dollar compensation packages—sometimes exceeding $10 million annually with equity—to lure elite AI researchers from academia, startups, and competitors. The recruitment push fed into a new experimental unit called TBD Lab, which has been instrumental in laying the groundwork for Avocado by experimenting with hybrid training techniques and diverse data pipelines.
Market reactions were swift and telling following the Bloomberg disclosure. Meta’s shares experienced a modest decline of 1.2% in after-hours trading, reflecting investor concerns over the costs and risks of this ambitious pivot away from a cost-effective open-source model. Conversely, Alibaba’s stock surged by 2%, buoyed by revelations that Meta’s TBD Lab incorporated Alibaba’s powerful Qwen series—particularly advanced iterations like Qwen 2.5—into Avocado’s training regimen alongside other third-party contributions such as Google’s Gemma and open variants from OpenAI. This marks a pragmatic, if surprising, embrace of Chinese AI technology by Zuckerberg, who had previously voiced apprehensions about potential state-influenced biases or censorship in models from firms like Alibaba. The integration leverages Qwen’s strengths in multilingual processing and cost-efficient scaling, helping Meta bolster Avocado’s capabilities without solely relying on its own datasets. Such cross-pollination highlights the interconnected nature of global AI development, even amid geopolitical tensions, and positions Avocado as a potentially more robust contender from the outset.
Internal Tensions Escalate in AI Leadership
Deep within Meta’s newly established Superintelligence Labs—the umbrella organization now overseeing all frontier AI pursuits—frictions are mounting as the shift to a closed-model strategy reshapes team dynamics and priorities. At the helm sits 28-year-old prodigy Alexandr Wang, founder of Scale AI, who joined Meta in June 2025 as Chief AI Officer through a landmark $14.3 billion investment deal that effectively acquihired much of his company and granted Meta a controlling 49% stake. Wang, known for his expertise in data labeling and evaluation pipelines critical to AI training, brings a philosophy favoring proprietary systems honed during Scale’s growth into a multi-billion-dollar enterprise serving clients like OpenAI and the U.S. Department of Defense. However, reports indicate growing frustration from Wang over Zuckerberg’s intensely hands-on micromanagement of AI roadmaps, resource allocations, and even model architecture decisions—a style that has reportedly led to heated internal debates and delayed iterations during Avocado’s development.
This clash of visions between a young disruptor and a veteran CEO illustrates broader challenges in scaling superintelligence efforts, where blending entrepreneurial agility with corporate oversight proves tricky. Compounding these issues, the reorganization has triggered high-profile exits, most notably that of Yann LeCun, Meta’s chief AI scientist for over a decade and a Turing Award winner revered as one of the “godfathers of deep learning” for pioneering convolutional neural networks. LeCun, who established Meta’s FAIR (Fundamental AI Research) lab in 2013, announced in November 2025 his departure by year’s end to found an independent startup centered on “world models”—AI systems capable of simulating and predicting real-world physics and environments with human-like intuition.
Sources attribute his exit partly to discomfort with reporting under Wang’s leadership in the revamped structure, after years of operating with significant autonomy. LeCun’s departure represents a loss of institutional knowledge and prestige for Meta, as his work influenced everything from computer vision to generative AI, and it signals potential brain drain risks as top talent weighs opportunities in a hot startup market flush with venture capital.
Launch Timeline Slips Amid Massive Investments
What was initially targeted for a late 2025 debut, Avocado’s rollout has been deferred to the first quarter of 2026, primarily due to persistent challenges in training stability, compute optimization, and achieving breakthrough performance on multimodal tasks like video understanding and long-context reasoning. Engineers have grappled with issues such as model collapse during extended fine-tuning and inefficiencies in distilling knowledge from massive foundation models, necessitating additional rounds of hyperparameter tuning and synthetic data generation. Despite these setbacks, a Meta spokesperson assured CNBC that “model training efforts are proceeding as planned and there have been no significant changes to our timeline,” framing the delay as a refinement rather than a derailment—a common tactic in the opaque world of AI announcements to manage expectations.
To underwrite this high-stakes race against OpenAI’s o1 series and Google’s upcoming Gemini 2.0, Meta has dramatically escalated its financial commitments. The company recently raised its 2025 capital expenditure guidance to $70 billion to $72 billion, up from prior estimates, with the bulk allocated to AI infrastructure including next-generation GPUs from Nvidia, custom silicon development through Meta’s MTIA chips, and sprawling data center expansions across the U.S., Europe, and Asia. This spending spree—among the largest in tech history—covers not just hardware procurement but also energy-efficient cooling systems, renewable power contracts, and global fiber optic networks to minimize latency.
It positions Meta to train models at unprecedented scales, potentially exceeding 10 trillion parameters for Avocado, while navigating supply chain bottlenecks and regulatory scrutiny over energy consumption. Ultimately, this investment gamble reflects Zuckerberg’s conviction that controlling next-generation AI will redefine Meta’s core businesses in social media, advertising, and emerging metaverse applications, ensuring long-term leadership in an industry where compute power increasingly dictates innovation speed.
