Summary: The Verge reported on March 12 that Meta moved Avocado, its next flagship model, from March to no earlier than May because internal benchmarks showed it trailing leading competitors. Reporting placed its capability between Google Gemini 2.5 and Gemini 3, while CNBC had reported in December 2025 that Avocado might become proprietary. The article urges teams relying on Meta models to keep routing flexible.
A New York Times report picked up by The Verge on March 12 says Meta has moved the launch of Avocado — the codename for its next flagship model — off March and to no earlier than May. The stated reason is blunt: internal benchmarks have the model running behind the field’s strongest competitors, and the reporting slots its capability between Google’s Gemini 2.5 and Gemini 3.
A second thread gives the slip extra weight. As far back as December 2025, CNBC reported that Avocado might arrive as a proprietary model, breaking with the open-source lineage of Llama. With timing and direction both wobbling, Meta’s model program has reached its hardest-to-read moment since the Llama era opened.
What the Internal Scores Revealed
In this industry, a slipped launch barely registers as news. What carries real information is the quoted line about “internal benchmarks showed it trailing.” It reveals how Meta measures itself: against the field’s strongest, never against its own previous generation. That yardstick is the correct one, and it is also the punishing one.
Being slotted between Gemini 2.5 and Gemini 3 is the more troublesome half, because the target being chased refuses to stand still. By the time Avocado surfaces in May, rival flagships may have moved a generation further. Waiting is how Meta avoids taking the field at a disadvantage, yet every additional month lifts the bar all over again.
A disclosure cost rides along too. The slip became public with its benchmark context attached, and that context says the same thing to every enterprise buyer weighing a Meta deployment as it does to Google and OpenAI: the coming successor to Llama is not yet able to hold the ground, and the market has heard it.
Open Weights at the Fork
Should the December reporting prove out and Avocado turn proprietary, the blast radius would not stop at Meta. For years, the Llama strategy cast Meta as the load-bearing member of the open-weights camp. Vacate that seat, and teams hunting open-source flagships lean harder on other suppliers — no small list in early 2026, when Moonshot’s Kimi K2.5 arrived under the MIT license and Zhipu’s GLM-5 shipped as an open-source flagship in its own right. Open source would survive Meta’s exit; the distribution of weight across it would not stay the same.
Downstream teams get a warning out of the same story: bolt your product to a single vendor’s open-source roadmap and you have handed that vendor’s business judgment the pen on your own strategy. The practical takeaway for anyone running open-weights infrastructure is to count Meta as one supplier among several, watch every supplier’s maintenance calendar, and drill the migration path before anyone needs it.
The Gap Before May
None of the rivals sat still across Meta’s two extra months of baking. Just the first quarter’s shipped items:
- Anthropic: Claude Opus 4.6 went live, bringing agent teams and a 1M-token context window (February 5)
- Google: Gemini 3.1 Pro landed in preview, doubling reasoning at unchanged pricing (February 19)
- OpenAI: GPT-5.4 arrived with professional work as its target audience (March 5)
Whatever Avocado meets in May will be a taller order. Two months gives a rival room to land a flagship launch, a pricing move, or a bundle of enterprise features — and any one of those rewrites the comparison Avocado has to win. Meta is not tuning its model in private against a fixed exam; it is tuning against a scoreboard the rest of the industry keeps updating in public. Teams built on Meta’s open models gain nothing from waiting. The useful move is keeping model routing flexible enough that, however Avocado lands — open weights or not, May or later — switching stays a configuration change rather than a rewrite.