Meta's internal 'Watermelon' model reportedly hits GPT-5.5-class performance

Meta's newest internal AI model, codenamed 'Watermelon,' reportedly reaches GPT-5.5-class performance on internal reasoning and coding benchmarks — a substantial jump over prior Llama-family models. Reports attribute the leap to vastly increased training compute and a unique proprietary dataset, and it arrives as Meta AI chief Alexandr Wang publicly claimed the company is 'finally catching up' to OpenAI, Google and Anthropic after a period of perceived stagnation.

The claim should be read against Meta's aggressive infrastructure and consumer-AI push: a reported $6.5B Samsung foundry deal for next-gen MTIA silicon, a 5-gigawatt compute goal by 2030, the 'Meta Compute' cloud plans, and consumer apps like Pockets and Vibes. Watermelon is the model layer underneath that strategy — Meta needs a frontier-competitive base to make its cloud and app ambitions credible.

However, the sourcing here is thin (an AI-news aggregator citing 'internal benchmarks'), and 'GPT-5.5-class' is an unverified, self-reported comparison. Meta has a history of benchmark claims that faced scrutiny after release, and the open-source-vs-closed question looms: it's unclear whether Watermelon (or a Muse Spark successor) ships with open weights or as a hosted-only model, which would be a significant shift given developer expectations.

What to watch: an official Meta announcement with reproducible benchmarks, the model's licensing, and whether it powers Meta Compute's API tier. Until then, treat the performance parity as a claim, not a result.