Meta’s AI Course Correction Spurs Internal Tension as “Avocado” Model Faces Delay

Meta Platforms is reshaping its artificial intelligence roadmap, replacing last year’s public push for the open-source Llama language models with a costly quest to build a proprietary successor, codenamed “Avocado.” People familiar with the effort say the new model, once slated for release before the end of 2025, is now expected in the first quarter of 2026 while engineers run extensive performance tests.

The strategic pivot follows a series of high-profile hires, larger capital-spending targets and mixed results from earlier model launches. Chief Executive Mark Zuckerberg championed Llama in 2024, predicting it would become the industry’s most advanced open-source system. By Meta’s October 2025 earnings call, he mentioned Llama only in passing, underscoring the company’s shifting priorities.

Expensive talent drive

In June 2025 Meta spent $14.3 billion to bring in Scale AI founder Alexandr Wang and several senior researchers. Wang was named chief AI officer and now leads the elite TBD Lab, the internal group developing Avocado. Former GitHub chief executive Nat Friedman heads product and applied research within the newer Meta Superintelligence Labs (MSL), while Shengjia Zhao, a ChatGPT co-creator, joined to refine model-training practices.

The hires contributed to Meta raising its 2025 capital-expenditure forecast to $70 billion–$72 billion, up from the lower end of a previously disclosed range. Analysts at KeyBanc Capital Markets wrote in November that investors are looking for evidence the spending will translate into competitive AI products and stronger returns.

Pressure to match rivals

Competitors have released successive upgrades over recent months. Google unveiled Gemini 3, OpenAI announced further improvements to GPT-5, and Anthropic introduced Claude Opus 4.5. Market watchers say no single model leads across every benchmark, but continual launches raise expectations for Meta to present a comparable system.

Meta’s own Llama 4, released in April 2025, failed to capture broad developer interest. Internally, senior engineers criticized open-source licensing after China-based DeepSeek’s R1 model reused elements of Llama’s architecture. The episode accelerated debate over whether future Meta models should remain open. Several people familiar with current plans say Avocado will likely be proprietary, restricting access to its weights and software components.

Organizational upheaval

The uneven Llama 4 rollout prompted management changes. Chief Product Officer Chris Cox no longer oversees the generative-AI unit, while veteran engineering vice president Aparna Ramani was put in charge of allocating compute resources across MSL. Vishal Shah, previously responsible for metaverse products, became vice president of AI Products to bridge traditional social-media teams with the new research groups.

Despite an annual revenue base exceeding $160 billion, Meta’s core advertising business continues to grow more than 20% a year, thanks partly to AI-enhanced targeting. However, many of the new AI leaders have limited ad-tech experience, reflecting Zuckerberg’s belief that Meta must tackle broader AI breakthroughs to avoid irrelevance.

Imagem: Internet

Cultural clash and longer hours

Current and former employees describe a rift between long-standing Meta engineering processes—characterized by cross-functional reviews and extensive internal tooling—and the fast-iteration approach favored by Wang and Friedman. The newer leaders promote a “demo, don’t memo” philosophy, encouraging smaller teams to prototype quickly using external agent-based coding tools. Workweeks of 70 hours have become common, the sources say, amid periodic layoffs and restructurings. In October, about 600 roles were cut within MSL, affecting units such as the Fundamental AI Research team and contributing to the departure of chief AI scientist Yann LeCun, who plans to launch a startup.

Infrastructure and spending

Meta is revamping its compute strategy while building the massive Hyperion data center in Richland Parish, Louisiana, through a $27 billion joint venture with Blue Owl Capital. To bridge capacity gaps, the company is increasingly renting cloud GPUs from providers like CoreWeave and Oracle. The heavy demand flows through to suppliers such as Nvidia, whose chief executive Jensen Huang recently highlighted Meta among customers using its hardware; Nvidia’s latest quarter showed 62% year-over-year revenue growth, underscoring the broader AI spending boom (Nvidia corporate site).

Product experiments and market reception

While Avocado remains in training, Meta has rolled out interim AI features. Vibes, a short-video generator launched in September 2025, relies on models from Black Forest Labs and Midjourney. Downloads trail those of OpenAI’s Sora 2, and creators have criticized Vibes for lacking realistic lip-synched audio. Still, the feature has driven additional traffic to the stand-alone Meta AI app.

Meta’s stock performance in 2025 has lagged the broader technology sector and Alphabet in particular. Analysts attribute investor caution to unclear timelines for high-end AI products and uncertainty around monetization. Zuckerberg told analysts in October that MSL is “off to a strong start” and said he believes Meta now has the industry’s highest concentration of AI talent. He pledged to share further progress in the coming months.

For now, employees and shareholders await evidence that Avocado can close the perceived gap with rival models—and justify the billions of dollars Meta has committed to an evolving, and increasingly proprietary, AI strategy.

Crédito da imagem: Manuel Orbegozo | Reuters