Meta's Superintelligence Labs has delivered its first AI models for internal testing, CTO Andrew Bosworth announced at a World Economic Forum press briefing in Davos on January 21. The team, roughly six months into its work, produced results Bosworth called "very good," though he declined to name the specific models or provide technical details.
What we actually know
The announcement, first reported by Reuters, is thin on specifics. Bosworth confirmed delivery but not what was delivered. Media reports from December pointed to two projects under development: a text-focused model codenamed "Avocado" and a multimodal image-and-video model called "Mango." The Wall Street Journal reported both were targeting first-half 2026 releases, though Meta hasn't confirmed these names publicly.
"There's a tremendous amount of work to do post-training," Bosworth said, "to actually deliver the model in a way that's usable internally and by consumers."
That caveat matters. Internal delivery to engineering teams is a milestone, but it's not a product. The gap between a working model and something reliable enough for Meta's 3.2 billion users can be months or years.
The pressure behind the progress
This announcement comes after a bruising year for Meta's AI ambitions. Llama 4, released in April 2025, landed poorly. Developers criticized underwhelming performance and inconsistent results. Worse, Meta faced accusations of submitting a specially tuned "experimental" version to the LMArena benchmark that differed from the public release, a move that drew rebukes from the evaluation platform itself.
The fallout was significant. Chris Cox, Meta's chief product officer and a 20-year company veteran, lost oversight of the AI division. Yann LeCun, widely considered a godfather of deep learning, left his chief AI scientist role in November 2025 to start his own firm. LeCun reportedly clashed with the new structure that placed him under Alexandr Wang, whose background was in data labeling rather than model development.
Wang's expensive mandate
Zuckerberg responded to the Llama 4 debacle by spending aggressively. Meta invested $14.3 billion for a 49% stake in Scale AI and hired its 28-year-old founder Wang as Meta's first chief AI officer. The reorganization consolidated all AI teams under a new entity called Meta Superintelligence Labs, with Wang leading and former GitHub CEO Nat Friedman overseeing products.
Wang's hire raised eyebrows. Scale AI built data labeling infrastructure, not frontier models. Whether that expertise translates to leading a superintelligence effort remains the central question. Recent Financial Times reporting suggested tensions between Wang and Zuckerberg, with Wang telling associates he found the CEO's involvement suffocating.
The internal structure now includes four groups: TBD Lab (led by Wang, handling foundation models), FAIR (research), Products and Applied Research (under Friedman), and MSL Infra (infrastructure). The Avocado model is reportedly being built inside TBD Lab.
What Bosworth didn't say
Missing from Davos: any benchmark comparisons, architectural details, or timeline for external availability. That's standard practice for internal milestones but notable given how much Meta has invested in this reset. Bosworth's framing, that 2025 was "tremendously chaotic" but 2025's bets are "starting to show favorable returns," suggests Meta is managing expectations carefully.
He did offer one prediction: 2026 and 2027 would see consumer AI trends firm up, with models increasingly handling everyday questions. That's a modest claim compared to the "superintelligence" language in Zuckerberg's original announcement.
Meta plans to continue building out infrastructure, including the Prometheus data center in Ohio and a massive Louisiana facility called Hyperion. The company's capital expenditure guidance for 2025 reached $70-72 billion, much of it AI-related.
The first models are in. Whether they're any good remains Meta's internal secret for now.
