Meta's Watermelon large language model, the successor to Muse Spark from Meta Superintelligence Labs, remains in training as of early October 2026 after internal documents targeted an October release window. Alexandr Wang stated in a July town hall that the model, trained on roughly 10x the compute of its predecessor, had reached parity with OpenAI's GPT-5.5 on internal benchmarks, though specific evaluations were unnamed and unverified externally. No frontier model debuted at Meta Connect in September, and no public weights, API access, or Text Arena leaderboard appearance has occurred. Traders are watching for any late-October announcement that could shift competitive positioning against GPT-5.5/5.6 and Claude models, as Meta continues scaling infrastructure for agentic and coding capabilities.