Meta Delays ‘Avocado’ AI Model Launch to May or Later
Meta’s eagerly anticipated next-generation language model, code-named “Avocado,” has been pushed back to at least May from an earlier planned rollout after internal tests showed the system is not yet performing at the level company leaders expected. Meta said the additional time will allow engineers to close gaps identified during benchmarking and better align the model’s capabilities with the company’s product plans.
According to people familiar with the matter, Avocado’s internal benchmarks placed it between Google’s recent Gemini 2.5 and the newer Gemini 3 in several core areas such as reasoning, coding, and nuanced writing — a performance profile that Meta’s product teams judged insufficient for a broad launch. The delay reflects an industry-wide push to make emerging models more reliable before exposing them to tens of millions of users.
The setback comes as Meta is investing heavily in an expansive AI roadmap that includes custom chips, infrastructure buildout, and a series of model releases over the coming year. Internal planning materials suggest leadership is leaning toward a more measured rollout. Instead of pushing out a model that isn’t fully ready, the company appears focused on refining performance through additional testing and staged improvements.
At the same time, discussions inside the company have reportedly explored a short-term workaround: tapping into Google’s Gemini models for select applications if Avocado takes longer than expected to meet internal benchmarks. Licensing would be an unusual but pragmatic move in an industry where first-to-market pressure must be balanced against product quality and safety. Any licensing talks would require careful negotiation and review of integration paths and data handling practices.
For users and partners, the practical effect will likely be phased: internal pilots and limited tests may continue while broader availability is postponed. Analysts say the delay is unlikely to derail Meta’s long-term AI ambitions but could compress the company’s product timeline for features that rely on Avocado-class reasoning. Investors are already attentive to the company’s cost-to-benefit calculus as Meta expands AI staffing and infrastructure spending.
Meta’s public communications have framed the pause as part of iterative engineering rather than a fundamental rethink. The company is expected to continue refining Avocado’s training mix, prompting rigorous re-evaluation of safety guardrails, hallucination rates, and alignment tests before any public-facing deployment. Industry observers note that such additional rounds of internal benchmarking and red-teaming are becoming standard practice as large model makers try to reduce user-facing failures.
As with other high-profile model launches, the Avocado timeline will be watched not just for technical milestones but for what it signals about competitive dynamics between major AI players. While a May rollout is now the near-term target, sources stress that “May or later” captures the reality that model development remains an empirical process: results — not calendar dates — will ultimately determine when the system is ready for prime time.
