Meta's Next-Gen AI Model Avocado Delayed to May Over Performance Testing Issues
The New York Times reported on March 12 that Meta's next-gen closed-source AI model 'Avocado' has been delayed from Q1 to May or later due to performance testing difficulties. Internal tensions under new CAIO Alexandr Wang, the pivot from open-source Llama to closed models, and a 2.55% stock drop paint a challenging picture for Meta's AI ambitions.
According to a New York Times report on March 12, Meta's next-generation AI model 'Avocado' has been delayed from its original Q1 2026 (March) target to May or later. The core issue is difficulties encountered during performance testing in training. Meta's stock dropped 2.55% immediately after the report, reflecting swift market reaction.
Avocado is Meta's first frontier model built under a closed-source strategy, abandoning the open-source approach of its Llama series. As a text-based LLM specializing in coding and reasoning, it carries Meta's ambitions to compete directly with OpenAI's GPT-5, Google's Gemini 3, and Anthropic.
What's Been Delayed: Key Takeaways from the NYT Report
The NYT report is straightforward: Avocado has run into difficulties with performance testing during training, pushing its launch timeline from the original March target to May or beyond.
Avocado is a coding and reasoning-focused LLM being developed under Meta Superintelligence Labs (MSL). A companion model called 'Mango' specializes in image and video generation. Both models are set to launch as closed-source, unlike the Llama series.
Meta officially pushed back on the report, stating that 'model training efforts are on track and there are no meaningful schedule changes.' However, the market didn't buy the explanation: shares dropped 2.55% on the day of the report.
Alexandr Wang and Internal Tensions at Meta AI
The delay isn't just about technical hurdles. The NYT report also highlights friction between new CAIO Alexandr Wang and Meta's existing culture.
Wang, formerly CEO of Scale AI, was recruited by Zuckerberg in a deal reportedly worth $14.3 billion. He took charge of Meta's entire AI organization with the creation of Meta Superintelligence Labs (MSL) in 2025, but the transition hasn't been smooth.
Since MSL's launch, 600 employees were laid off, and Yann LeCun, long the face of Meta AI as chief AI scientist, departed. An externally recruited leader restructuring at scale while reshaping company culture inevitably created internal friction. Analysts suggest this organizational instability has contributed to delays in Avocado's development timeline.
From Open Source to Closed: Why the Strategic Pivot
One of the biggest reasons Avocado commands attention is that Meta's AI strategy itself is undergoing a fundamental shift. The company that championed 'AI democratization' by open-sourcing its Llama series is now going proprietary starting with Avocado.
The direct catalyst was China's DeepSeek. When DeepSeek leveraged Meta's Llama architecture to rapidly build competitive models, Meta's leadership grew frustrated that their massive R&D investments were effectively subsidizing competitors.
The underwhelming reception of Llama 4, launched in April 2025, further accelerated the pivot. When the open-source strategy failed to translate into competitive advantage, Meta shifted toward the closed-model, direct-monetization approach of OpenAI and Google.
Competitive Landscape and Market Reaction
Avocado's delay isn't just Meta's problem: it illustrates the intensity of competition across the entire AI industry. OpenAI is preparing GPT-5, Google is working on Gemini 3, and Anthropic has its next-generation Claude in the pipeline, while Meta's first closed-source model stumbles before launch.
The market response was harsh. Meta stock fell 2.55% immediately following the NYT report. With Meta having committed $115–135 billion in 2026 AI capital expenditure, a 73% increase year-over-year — a delay in the first deliverable reads as a warning signal to investors.
That said, a leaked internal memo from February praised Avocado as 'the most capable pre-trained model in Meta's history,' so it's premature to write off the model's potential entirely. The question is when and how that potential can be proven to the market.
Avocado's Report Card Will Define Meta AI's Future
Avocado's delay carries significance beyond a simple schedule slip. Meta's first real test under its new paradigm, abandoning Llama's open-source strategy, overhauling the organization under new leadership, and committing over $100 billion in investment, is showing cracks.
The key question is whether Avocado, when it finally launches after May, can validate the internal memo's claim of 'SOTA-level performance from pre-training alone.' With OpenAI, Google, and Anthropic racing ahead, how quickly Meta can close the gap remains to be seen. Avocado's final report card will provide the answer.
