Meta's 'Watermelon' Model Reportedly Matches GPT-5.5 Benchmarks

Summary: Meta Superintelligence Labs head Alexandr Wang announced at a July 3 internal town hall that Watermelon—Meta's next flagship model trained with roughly 10× the compute of Muse Spark—has achieved parity with OpenAI's GPT-5.5 on several benchmarks.

Key Points

• Watermelon reportedly matched GPT-5.5's MMLU score of 92.4% and edged it on HumanEval (96.3% vs. 96.1%) per internal figures
• Uses ~10× more compute than Meta's previous generation model Muse Spark; no public release date announced
• Announcement came amid internal debate over excess compute capacity and whether AI progress has slowed
• Wang's claim is unverified—no reproducible published evaluation has been released; treat as a directional signal

Why It Matters

If Watermelon delivers frontier-class performance when it eventually ships as an open-weights model, any developer worldwide could run it for free, reshaping the open vs. closed model dynamic yet again. The caveat is real: internal benchmarks at one company's town hall are not the same as a published, independently verified result.

Read More

• Meta's 'Watermelon' Matches OpenAI GPT-5.5 Benchmarks — Benzinga
• Meta's Watermelon Matches GPT-5.5 Benchmarks — Let's Data Science