Meta's Avocado Model Slips Again: Inside the Llama 5 Closed-Source Pivot

Meta's next-generation flagship model, codenamed Avocado, has slipped past its planned May 2026 launch window — the third delay in roughly nine months. The slip is the clearest sign yet that Meta Superintelligence Labs' pivot from open-source Llama to a closed frontier model is harder than Mark Zuckerberg expected.

The model, widely referred to internally as Llama 5, completed pre-training in January and was originally scheduled for a March release before slipping first to May and now into the summer.

What Avocado Is Supposed to Be

Avocado is being developed inside Meta's elite TBD Lab — the unit led by Alexandr Wang, the Scale AI founder Meta acquired in June 2025 for a reported $14.3 billion package. Internal memos describe Avocado as 10x more compute-efficient than Llama 4 Maverick on text tasks and over 100x more efficient than the never-shipped Llama 4 Behemoth.

The architectural details suggest a mixture-of-experts design with aggressive routing and a new tokenizer optimized for multilingual workloads. Meta has been quietly hiring Chinese open-source researchers and reportedly studied Alibaba's Qwen training pipeline as a reference point.

The strategic shift is more important than the architecture. Unlike every prior Llama, Avocado will be closed-source. Meta cited concerns about Chinese labs — particularly DeepSeek — directly building on the open Llama lineage. The company also concluded that the developer goodwill from open weights was not converting into commercial advantage at the rate Zuckerberg wanted.

Why the Delay Matters

Internal testing has reportedly shown Avocado falling short of Google's Gemini 2.5 and 3.0, and meaningfully behind Anthropic's Claude Opus 4.7 and OpenAI's GPT-5.5, on reasoning, coding, and writing benchmarks. The compute-efficiency gains do not appear to fully translate to capability gains, and that has spooked the launch team.

The Trending Topics EU report goes further: Meta has reportedly considered licensing Google's Gemini as a backstop product, a scenario that would have been unthinkable two years ago when Meta was the open-source AI standard-bearer.

Why this matters: Meta is the only top-five AI lab without a current commercially shipping frontier model. The longer Avocado slips, the more enterprise and consumer attention flows to Anthropic, OpenAI, and Google — and the harder it becomes to recapture mindshare with a delayed launch.

The End of Meta's Open-Source Era

Meta has been the dominant force in open-source AI since Llama 2 in 2023. Every major derivative ecosystem — Mistral, Llama-based fine-tunes, the entire open-weights coding model wave from Chinese labs — traces back to weights Meta released for free.

That era is ending. Buttondown's analysis framed Avocado as "the end of an open-source era," and the broader industry consensus is now that Meta's strategic pivot is permanent. Smaller open-weights models from Meta may still ship as marketing exercises, but the frontier model is moving behind closed doors.

For the open-source community, the loss matters. Without a heavyweight open-source benchmark to anchor the ecosystem, the Chinese labs — Z.ai, MiniMax, Moonshot, DeepSeek — become the de facto leaders, all four of which shipped competitive open-weights coding models in a 12-day window last month.

The Yann LeCun Departure

Avocado's troubles compound an earlier blow. Yann LeCun, who founded Meta's Fundamental AI Research lab in 2013 and won the Turing Award for his work on deep learning, resigned in late 2025 after an October 2025 layoff wave targeting FAIR. LeCun's stated reason was "strategic differences" with the new closed-source direction.

LeCun has since founded AMI Labs in Paris, raised $1.03 billion at a $3.5 billion pre-money valuation in March — Europe's largest seed round ever — and is building "world models" based on his JEPA architecture rather than transformer-based language models.

The departure left FAIR diminished and shifted the center of gravity at Meta to Wang's TBD Lab. Inside Meta, the joke for months has been that Avocado is "Alex's model, not Yann's." That cultural shift is now reflected in product timing.

Industry Implications

For OpenAI and Anthropic, the longer Avocado slips, the longer they extend their enterprise lead. PwC, Blackstone, Goldman Sachs, and other $1M+ Anthropic accounts are not waiting on Meta — they are deploying Claude today. For Google, Meta's pivot makes Gemini's distribution advantage even more important: if Meta is no longer an open-source counterweight, Google's enterprise positioning improves.

For developers, the practical loss is the Llama community ecosystem. Hundreds of fine-tunes, derivatives, and inference optimizations have been built on Llama weights. Without a Llama 5, that work either stalls or migrates to Chinese open-weights alternatives like GLM-5.1, Kimi K2.6, or DeepSeek V4.

Expert Perspectives

Reactions among AI researchers on X have ranged from disappointment to frustration. The most-discussed thread argues that Meta squandered a unique market position — the only major lab credibly committed to open weights at the frontier — for a closed-source strategy that has not yet produced a competitive model.

Wang's defenders point to the compute-efficiency claims and argue that a 10x improvement in training cost is the kind of structural advantage that takes a year to convert into shipped product. The bull case is that Avocado lands in late summer, is genuinely competitive, and Meta resets the narrative.

What's Next

Three things to watch. First, a revised Avocado launch date — Meta has not committed publicly, but internal targets reportedly point to late Q3 2026. Second, whether Meta ships any open-weights model in the meantime to maintain developer relations. Third, the Gemini licensing rumor: if Meta strikes a deal with Google, the open-source pivot is effectively complete.

The Bottom Line

Avocado's delay is more than a missed launch window — it is the moment Meta's shift from open-source leader to closed-source competitor becomes visible. The longer the slip, the more ground OpenAI, Anthropic, and Google gain. The torch has already passed to Chinese labs, and Meta's return to the open-weights frontier looks increasingly unlikely.