Summary: Meta CTO Andrew Bosworth said at a Davos press briefing on January 21, 2026 that Meta Superintelligence Labs had completed internal handoff of its first key models about six months after formation. He called them very good but said substantial post-training work remained before wider or consumer availability. Meta executives also framed Ray-Ban glasses and neural bands as key input surfaces for superintelligence.
Meta CTO Andrew Bosworth took a press briefing held alongside the World Economic Forum in Davos on January 21, 2026, to confirm a milestone: Meta Superintelligence Labs — the unit Mark Zuckerberg created in his 2025 overhaul of Meta’s AI setup — has completed the internal handoff of its first key models this month, six months or so after the lab’s founding. He graded the models “very good” and credited them for having “shown significant promise,” declined to say which ones they were, and stressed that “there’s a tremendous amount of work to do” on the post-training front before anything reaches the wider company and, down the line, consumers.
The backdrop is what gives the announcement weight. While Google and OpenAI never stopped raising the bar, Llama 4 took heat for falling behind them; MSL, born from aggressive big-spending talent raids, is the catch-up answer. Models landing internally at the half-year mark stand as the first checkable proof that the restructuring is producing output at all.
Six Months to First Models: The Score and the Caveat
The timing fits the December reporting that had Meta carrying two codenames — a text model called Avocado, originally slated for a Q1 launch, and Mango, a multimodal model spanning images and video. Bosworth’s account squares with that schedule, though he would not say which models had landed, offering only that heavyweight results had been delivered. The caveat deserves equal billing: post-training is what turns a pretrained model into a usable product, and by his own telling “a tremendous amount of work” is still waiting in that phase.
Looking back, he called 2025 a “tremendously chaotic year” while insisting the wagers — power purchases, infrastructure, and a lab of its own — are already paying back in a favorable direction. Read it two ways at once: an account of progress, and a shield for the money already spent. The cash is out the door, models are beginning to show up, and the real reckoning comes later.
Where Superintelligence Lives: Glasses and Neural Bands
Across the same summit, the conversation at Meta’s Davos pavilion made the vehicle explicit. Derya Matras, the VP responsible for Europe, the Middle East and Africa, argued that wearables carry an advantage no phone or computer can match: “They see what you see, they hear what you hear” — and they read the environment around you. “That’s how we get closer to superintelligence.” The hardware she singled out was Ray-Ban smart glasses and neural bands, meant to give AI a grasp of each user’s dreams, interests, and goals — framed as “giving that power to billions of people.”
The operating numbers underwrite the story. Nicola Mendelsohn, who leads global business, put Meta’s family of apps at 3 billion daily active users and called AI “the backbone our recommendation systems.” Meta, meanwhile, is selling the Ray-Ban Display glasses — and in early January it put their international rollout on hold so that US orders could be filled first. In this narrative, wearables are no accessory line; they are the superintelligence pitch’s main input surface.
LeCun’s Exit and the Unsettled Route
Yann LeCun, once Meta’s chief scientist, dismissed LLMs as a “dead end” for anyone heading to superintelligence, and he left after Meta brought in Alexandr Wang, the Scale AI founder and an LLM believer, among other hires. Questioned on that history at Davos, Matras answered: “Yann has a particular view and he is a brilliant scientist,” adding that Meta now fields “perhaps the most talented team in the industry.”
That route dispute was never truly settled; reorganization simply papered over it. The wager inside MSL is LLM-centric scaling. Delivering internal models at six months shows how fast the team executes — but nothing in public yet proves this road gets to the destination.
2026–2027: The Two Years That Fix Consumer AI
On timing, Bosworth expects consumer AI usage patterns to settle across 2026 and 2027, because today’s models already manage “the kinds of things that you ask every day with your family, your kids,” while harder demands are still being worked out. Practically, the coming two years are the window Meta has to load its models into glasses and into its apps, and to claim everyday scenarios before habits set.
Worth reading against the year’s early signals: the January 2026 opening outlook already carried reports of a possible closed-weights turn for Meta’s flagship. Now that the first internal models exist, pressure is building on both fronts — when things ship, and whether this route is the right one.
Three Calls for Developers and Product Teams
Three working judgments. Leave Meta’s frontier models off near-term roadmaps — with the CTO himself repeating “tremendous amount of work” around post-training, no credible date exists for external availability. Wearables amount to a new surface for interaction: once input arrives as continuous sensing from glasses and wristbands, the defaults product design leans on — privacy, notifications, context management — have to be rewritten. And note the competitive signal in a reorganized lab putting internal models together within six months: it resets how fast a frontier team is expected to come together, and the window for catching up is narrower than most plans assume.