Meta's 'Avocado' Keeps Slipping: Closed-Source Gamble Stumbles as Company Eyes Gemini License

With Google I/O 2026 dominating the AI news cycle on Tuesday, Meta finds itself in an uncomfortable position: the company that once evangelized open-source AI to the world is now quietly shopping for a proprietary model license from one of its fiercest competitors, as its own flagship closed-source system falls further behind than its engineers expected.

A Year of Delays

Meta’s frontier model project, internally codenamed Avocado, has become one of the more awkward stories in the AI industry. Originally targeted for a Q4 2025 internal milestone, the model slipped to Q1 2026, then to March, then to May. May is now more than two-thirds gone with no announcement, and people familiar with the project say the most likely public release window is June — assuming the engineering team can close the performance gaps that have repeatedly pushed the timeline.

The delays are significant beyond mere scheduling. Meta set an ambitious internal goal: Avocado would be its first proprietary, closed-source frontier model, a direct successor to the Llama family but sold as a commercial product rather than released under an open license. It was supposed to demonstrate that Meta’s AI investment — the company has committed between $115 billion and $135 billion in capital expenditure for 2026 alone, with superintelligence listed as a stated goal — could produce a model competitive with OpenAI and Anthropic at the frontier.

The Performance Problem

Internal evaluations tell a complicated story. Avocado reliably beats Google’s Gemini 2.5 Pro, the model that preceded the I/O 2026 generation, on reasoning, coding, and writing benchmarks. That sounds promising until you account for the competitive landscape that has moved while Avocado was in development: Gemini 3.0 and 3.5, Claude Mythos, and GPT-5.5 have all shipped since Avocado’s original target date, raising the bar the model needs to clear to be commercially relevant.

According to people familiar with the internal evaluations, Avocado struggles most on complex reasoning chains — the kind of multi-step logical inference required for advanced coding assistance and scientific analysis. Long-horizon planning tasks, where the model needs to maintain coherent intent across dozens of reasoning steps, are reportedly the weakest area. These happen to be precisely the dimensions where Anthropic’s Mythos has established the widest lead over the rest of the field.

The irony runs deep: Meta’s strategy of building a proprietary model to compete with OpenAI and Anthropic appears to have underestimated how quickly those companies would themselves advance during the development window.

The Gemini Licensing Discussion

The detail that has attracted the most attention — and generated the most internal friction at Meta — is a reported discussion about licensing Google’s Gemini technology. According to people briefed on the talks, Meta AI leadership raised the idea of temporarily incorporating Gemini capabilities into Avocado or other Meta AI products to close the competitive gap while the company’s own model catches up.

No agreement has been reached, and it is unclear whether the talks progressed beyond preliminary exploration. But the very fact that the conversation happened illustrates the bind Meta is in. The company that built its AI credibility on democratizing access to powerful open weights models — Llama 2, Llama 3, and their derivatives have been downloaded hundreds of millions of times and power thousands of third-party applications — is now contemplating whether to depend on a rival’s proprietary infrastructure to remain competitive in the consumer AI market.

There is a practical logic to the idea: Google DeepMind has licensing arrangements with other companies, and a Gemini-powered Meta AI assistant would be meaningfully better than what Meta can ship today. But the reputational cost, combined with the strategic dependency it would create, has made the internal debate contentious.

Open Source to Closed Source: The Shift No One Expected

The deeper story here is what Avocado represents for Meta’s AI strategy. For years, Meta’s position was consistent and principled: open-source AI was good for the ecosystem, good for safety through transparency, and good for Meta because it created goodwill and third-party investment in the Llama ecosystem. Zuckerberg argued publicly that Meta should be the company that democratized frontier AI the way Android democratized mobile computing.

Avocado abandons that position. Unlike Llama, Avocado is designed as a commercial product with closed weights — users will pay to access it through Meta AI, and external developers will not be able to download and run it independently. The company has not fully explained the strategic reasoning behind the shift, but the commercial logic is apparent: as frontier models become genuinely revenue-generating products, Meta needs a proprietary asset it can monetize rather than giving away its most capable models.

The question is whether that shift is strategically sound given the current competitive dynamics. Meta has built a large and loyal developer community around the open Llama ecosystem. Avocado — closed, delayed, and reportedly underperforming expectations — risks alienating that community without delivering the competitive product that would justify the trade-off.

What Comes Next

Meta faces a narrowing window. Google I/O 2026’s announcements on Tuesday pushed the competitive bar higher: Gemini 3.5 Pro is now publicly available and Gemini 3.2 Flash is being deployed at scale across Google’s products. Every week Avocado remains in development, the model it is being benchmarked against improves.

The company still has significant assets: more than 3 billion daily active users across its family of apps, a vast social graph that could differentiate an AI assistant in ways pure model capability cannot, and an advertising infrastructure that could fund AI deployment at a scale few companies can match. But Avocado needs to exist and perform before those advantages can be activated.

A June launch, if it happens, would give Meta a product to show at its own potential developer events this summer. A further delay into Q3 would start to look less like an engineering decision and more like a strategic crisis.

For the broader AI industry, Meta’s difficulties are a data point in the ongoing debate about whether open-source development and frontier-model competition are compatible at this stage of the technology cycle. Right now, the evidence suggests that closing the gap with the best proprietary models requires concentrated engineering resources, focused research agendas, and training runs that dwarf what any open collaborative process has managed. Meta is learning that lesson at the scale of tens of billions of dollars.