Zuckerberg Hints Meta's Strongest Model Watermelon is Coming Soon
Meta CEO Mark Zuckerberg hinted in a video interview that the company's strongest AI model, Watermelon, is on the way, describing it as a significantly more advanced pre-trained model that feels good and will be launched soon.
The model, internally codenamed Watermelon, is the successor to Muse Spark (formerly known as Avocado). Alexandr Wang, head of Meta Superintelligence Labs, stated at a July employee meeting that it is still in training, using about 10 times the computing power, and has caught up with OpenAI's GPT-5.5 on internal benchmarks, although no specific benchmark names or scores have been disclosed.
Zuckerberg previously teased it with a watermelon emoji and confirmed it as the next major model in forums like the Sources Podcast. Meta aims to release it in October, but as of now, no official weight, model card, price, or exact date has been announced, and it did not appear at Meta Connect.
Wang also previewed an upcoming update for Muse Spark, with significant improvements in coding and agent capabilities, and mentioned a programming model nearing the level of Claude Opus will be released soon; Zuckerberg acknowledged that the progress of AI agents has not accelerated as expected.
Meta's capital expenditure for 2026 is expected to be between $125 billion and $145 billion for chips and data centers, with Watermelon representing a large-scale investment in computing power to catch up with the forefront, while the company advances the Muse series open-source weights and consumer-grade agent products like Hatch.
The event is driven by internal benchmark claims and the CEO's public announcements, with funding directed towards large-scale training power and Superintelligence Labs, benefiting investors holding Meta stock and anticipating AI monetization, while the competitive landscape faces pressure to validate independent benchmarks and actual product delivery.
The model has not yet been publicly released, and performance claims rely on internal assessments, with the market awaiting formal benchmarks and release confirmations.
Source: Public Information
ABAB AI Insight
Mark Zuckerberg appointed Alexandr Wang to lead Meta Superintelligence Labs in 2025, pushing the company from a core advertising model towards cutting-edge models and personal superintelligence. The historical path includes the April 2026 release of Muse Spark, followed by rapid iterations Muse Spark 1.1/1.2 and coding agents, while maintaining some open-source strategies, but the commercialization of agents is slower than expected.
On the capital front, the company has raised its annual capital expenditure to $125-145 billion for chips and data centers, reallocating resources from traditional social advertising to 10 times the computing power for training Watermelon to achieve parity with the internal benchmark of GPT-5.5, motivated by narrowing the gap with OpenAI and Anthropic and transforming the model into consumer agents and API revenue, rather than just selling computing power.
Similar cases include OpenAI's intensive pursuit from the GPT-4 to GPT-5 series and Google DeepMind's ongoing scale expansion. Currently, Meta is in a phase of expanding from a narrative of catching up to delivering validation, relying on internal claims and CEO announcements to maintain market expectations.
Essentially, this is about technological substitution and capital concentration: competition for cutting-edge models has shifted from algorithmic innovation to the scale of computing power and training investment. The mechanism is based on the path dependency of '10 times the computing power for benchmark parity,' allowing leaders to lock in performance gaps through continuous capital expenditure, while the absence of independent validation and release delays temporarily retains pricing power within the company's narrative rather than in public benchmarks.
ABAB News · Cognitive Laws
10 times the computing power for one benchmark parity
Teasing is a signal, release is validation
Internal benchmarks are stories, external benchmarks are prices