Meta’s next frontier large language model, internally codenamed Watermelon, remains in training following a reported July pre-training pause that shifted timelines. Chief AI officer Alexandr Wang told employees in early July that the model, running on roughly 10× the compute of the April-released Muse Spark series, had reached parity with OpenAI’s GPT-5.5 on internal benchmarks, though no specific tests or external evaluations were disclosed. Reports from The Information in August cited internal documents targeting an October 2026 public release, with the model positioned to power Meta AI agents and platforms. No launch occurred at the September Meta Connect event despite a watermelon emoji tease from Mark Zuckerberg, leaving trader focus on whether verification, safety checks, or further scaling will push the date later.