Meta’s new AI unit rolls out first key models internally

Meta Platforms’ new AI lab has delivered its first high-profile AI models internally this month, said CTO Andrew Bosworth.

Bosworth made the statement during a briefing at the World Economic Forum in Davos, calling the models created by Meta Superintelligence Labs “very good.”

Media outlets reported in December that Meta was developing a text-based AI model codenamed Avocado for a first-quarter launch, alongside an image- and video-focused model known as Mango.

Bosworth did not specify which models had been rolled out internally.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

The strategic significance of Meta’s new models remains to be seen

The performance of new internal models from Meta Superintelligence Labs should be compared with rivals such as OpenAI’s GPT-5 updates and Google’s Gemini 3 1.
Extra scrutiny is warranted because recent setbacks have been faced with Meta’s Llama models 2.
The intended internal uses should be spelled out, including whether engineering teams are being equipped to speed up coding for a 2026 goal, or whether new consumer features across Meta’s apps are being targeted 3.
The release approach should be confirmed, including whether an open-source path like Llama 3.1 will be followed or whether a proprietary route like the rumored ‘Avocado’ model will be taken 1.

Meta’s strategic ambiguity creates a dual opportunity for the AI ecosystem

MLOps (machine learning operations) and cloud infrastructure providers could see two paths if Meta shifts from open source to proprietary models 1.
If Meta ships another open-source release, vendors can sell hosting, fine-tuning (adapting a model to a specific business use), plus support for enterprise adoption built around Llama 3.1 4.
If the new models stay closed and arrive through an API (application programming interface), tooling startups can build abstraction layers plus cost-control software.
That setup can help enterprise software teams reduce vendor lock-in with a single closed provider such as OpenAI or Anthropic, while keeping a multi-model strategy 4.