Meta Develops Next-Gen Large Model 'Watermelon' with Increased Computing Investment

Meta is developing a new flagship large model codenamed 'Watermelon' to catch up with OpenAI and Anthropic. According to an exclusive report by Business Insider, Meta's head of superintelligence, Alexander Wang, revealed internally that the model's training compute is an order of magnitude higher than its predecessor Avocado, and it has matched OpenAI's GPT-5.5 on multiple benchmarks. Meanwhile, Meta recently announced the sale of idle AI computing power, sparking market concerns about overcapacity, but later signed a contract with Samsung for over 10 trillion won (approximately $1 billion) to produce 2nm AI chips, indicating continued investment in AI infrastructure.

Model Progress and Performance

• Watermelon is the successor to Avocado and Muse Spark, currently in the training phase.
• Computing resources are an order of magnitude higher than the previous generation, significantly enhancing logical reasoning and content generation capabilities.
• On mainstream benchmarks, performance is on par with OpenAI GPT-5.5, placing it in the industry's top tier.

Computing Power Sales and Market Reaction

• On July 1, 2025, Bloomberg reported that Meta plans to sell idle AI computing power and model access through the 'Meta Compute' project, similar to AWS Bedrock and CoreWeave models.
• Following the announcement, Meta's stock initially rose then fell, dropping 4.9% on July 2; the US semiconductor index fell 5.44%, with Micron and SanDisk dropping over 10%.
• The market feared a shift from AI infrastructure shortage to temporary oversupply, but Meta's subsequent actions indicate continued investment.

Samsung Foundry Order

• According to Caixin, Samsung and Meta signed a contract worth over 10 trillion won (approximately $1 billion) for AI chip foundry services, using 2nm process to mass-produce hundreds of thousands of chips.
• This order shows that Meta has not reduced computing investment but is optimizing the utilization of existing assets.

Impact and Interpretation

• Meta's dual strategy: selling older inference GPUs to improve asset utilization while massively purchasing advanced chips to train next-generation models.
• Market panic over computing overcapacity may be exaggerated; high-end training compute remains scarce.
• Netizens joked that Meta's model naming from Avocado to Watermelon increases 'water content,' predicting the next generation might be Coconut.