Meta’s superintelligence chief Alexandr Wang told staff in early July that Watermelon, Meta’s next flagship LLM, has already caught up to OpenAI’s GPT-5.5 on “key benchmarks.” Watermelon is still in training. There’s no release date, no API, and — conveniently — no word on which benchmarks.
What Watermelon actually is
A frontier text model, successor to Muse Spark (codenamed Avocado), the first model out of Meta Superintelligence Labs back in April. Watermelon burns an order of magnitude more compute than Spark, on a stack Wang says was rebuilt from the ground up. The problem it’s meant to solve is simple and existential for Meta: stop trailing OpenAI and Google at the frontier after a $14B talent spree.
Why it’s worth watching
The real edge isn’t compute — anyone with $130B in capex has that. It’s the training data: Facebook, Instagram, WhatsApp, Threads. Nobody else can copy that corpus. A single-sourced town-hall benchmark isn’t a reproducible eval, so treat “caught up” as a signal, not a fact. But the signal is loud: Meta is finally showing up to the frontier race.
You Might Also Like
- Meta Muse Spark 14b and 9 Months Later Alexandr Wang Delivers but its Closed Source
- Meta Hatch Metas Openclaw Rival Running on Claude Inside Instagram and Whatsapp
- Gpt oss 120b Openai Finally Goes Open Source and its Worth the Wait
- Meta Just Cracked Open Whatsapp for Third Party ai Chatbots but Theres a Catch
- Cognee Raises 7 5m Backed by Openai and Meta ai Founders can Knowledge Graphs fix Rags 40 Failure Rate
Top AI Product 
Leave a comment