On March 12, The Verge relayed a New York Times report: Meta has pushed its next flagship model, codenamed Avocado, from March to at least May. The reason was blunt — internal benchmarks show it trailing the top competitors, with reporting positioning its performance between Google’s Gemini 2.5 and Gemini 3.
A second thread makes the slip heavier. CNBC reported back in December 2025 that Avocado may ship as a proprietary model rather than continue the Llama open-source line. With both the schedule and the strategy in flux, Meta’s model program is at its most uncertain point since the Llama era began.
What the Benchmarks Said
A two-month delay is ordinary industry news. The informative part is the phrase “internal benchmarks showed it trailing.” Meta is grading itself against the strongest rivals, not against its own previous generation. That is the right standard, and it is also the brutal one.
The positioning between Gemini 2.5 and Gemini 3 is the deeper problem: the target is moving. By the time Avocado arrives in May, the competitors’ flagships may have advanced another generation. Delaying avoids launching at a disadvantage, but every extra month raises the passing grade again.
There is also a disclosure cost. The delay surfaced with the benchmark context attached, which tells every enterprise buyer weighing a Meta deployment the same thing it tells Google and OpenAI: the next Llama-successor is not ready to defend the position, and the market now knows it.
A Crossroads for Open Weights
If the December reporting holds and Avocado goes proprietary, the impact reaches beyond Meta. The Llama strategy made Meta the pillar of the open-weights camp; if that seat empties, the open-source flagship options tilt further toward other suppliers — and early 2026 was not quiet on that front, with Moonshot’s Kimi K2.5 released under MIT and Zhipu’s GLM-5 shipping as an open-source flagship. Meta stepping back would not kill open source, but it would shift the center of gravity.
For downstream teams it is also a reminder: tying your product to one vendor’s open-source roadmap means outsourcing a strategic decision to someone else’s business judgment. The practical reading for anyone running open-weights infrastructure is to treat Meta as one supplier among several, keep an eye on each one’s maintenance calendar, and rehearse a migration path before it is needed.
The Window While Meta Waits
The competitors did not pause during Meta’s two months of extra baking. The first quarter’s shipped items alone:
- Anthropic: Claude Opus 4.6 with agent teams and a 1M-token context window (February 5)
- Google: Gemini 3.1 Pro in preview, 2x reasoning at the same price (February 19)
- OpenAI: GPT-5.4, aimed at professional work (March 5)
By May, the bar Meta has to clear will only be higher. Two months is enough time for a rival to land a flagship launch, a pricing move, or an enterprise feature bundle — any of which resets the comparison Avocado has to win. Meta is not just improving its model in private; it is improving it against a scoreboard that keeps updating in public. Teams that depend on Meta’s open models should not wait; they should keep model routing flexible enough that whichever way Avocado lands — open or not, May or later — the switch is a configuration change, not a rewrite.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.