Meta Platforms’ recently formed artificial intelligence research unit has produced its first high-profile AI models for internal use, according to the company’s chief technology officer.
Early progress at Meta Superintelligence Labs
Speaking at a press briefing on the sidelines of the World Economic Forum in Davos on Wednesday, Meta CTO Andrew Bosworth said the models developed by the Meta Superintelligence Labs team were showing strong early results. The lab was established last year as part of a broader overhaul of the company’s AI strategy.
Bosworth noted that the team has been working for around six months and is still in the early stages of development. Despite that, he described the internally delivered models as “very good” and said they demonstrated significant promise.
Speculation around upcoming models
Bosworth did not confirm which specific models had been delivered internally. However, media reports in December suggested Meta was developing several new systems, including a text-based model reportedly codenamed Avocado, as well as an image and video-focused model known as Mango. Those reports indicated a potential launch timeline in the first quarter of this year.
Meta has not publicly detailed the capabilities or intended release plans for these models.
Leadership shake-up and competitive pressure
Meta’s AI efforts are being closely watched following major changes led by chief executive Mark Zuckerberg. Over the past year, the company has restructured its AI leadership, created the new lab and aggressively recruited top researchers with lucrative compensation packages. The moves signal Meta’s ambition to compete at the highest level in a crowded and fast-moving AI landscape.
The company has faced criticism over the performance of its Llama 4 model, particularly as rivals such as Google and other major technology firms have gained momentum with increasingly capable and commercially attractive AI systems.
Focus beyond training
Bosworth emphasised that delivering a usable AI product involves much more than training a model. He said there is extensive work required after training to make systems reliable, safe and practical for internal teams and eventual consumer use.
(With inputs from Reuters)