Social media giant Meta Platforms (META) was recently revealed to be accelerating the development of its next-generation artificial intelligence models. The company plans to release new-generation image/video models and large language models in the first half of 2026 and is actively exploring cutting-edge “world model” technology. These series of moves reflect the company’s continued push in the AI competition and a potential strategic shift from its previously emphasized open-source model toward pursuing cutting-edge technologies, possibly adopting a proprietary model approach.
According to reports, Meta is simultaneously developing two core AI models. One is a new-generation image and video AI generation model codenamed “Mango,” aimed at enhancing the company’s competitiveness in the field of generative visual content. The other is a next-generation large language model codenamed “Avocado,” with a key focus on improving professional capabilities like programming. Image generation has become a crucial battleground for AI giants, from Google’s Nano Banana to OpenAI’s Sora, and Meta’s previously launched Vibes video generator. All major players are actively positioning themselves, with OpenAI’s CEO also emphasizing that AI image generation is a “sticky” feature for attracting users.
This development plan further indicates a potential adjustment in Meta’s AI strategy. Previously, Meta championed the open-source Llama series models. However, recent reports suggest the Avocado model, expected to be released in the first quarter of 2026, might adopt a proprietary rather than open-source model. Catalysts for this shift may include the less-than-expected market response following the release of Llama 4 and potential intellectual property risks associated with the open-source strategy. To support cutting-edge R&D, Meta has undergone team restructuring, hiring Scale AI founder Alexandr Wang to lead a newly formed superintelligence department. It has also recruited a significant number of researchers from companies like OpenAI and substantially increased its capital expenditures.
Beyond specific application models, Meta executives have revealed the company is in the early stages of researching and developing “world models.” A world model is an AI technology designed to learn and understand the environment by absorbing information such as visuals, inspired by the human way of perceiving the physical world. This technology is believed to equip artificial intelligence with the ability to understand the real three-dimensional world, holding significant importance for achieving more advanced embodied intelligence. However, world models also face significant technical challenges, including immense computational power demands, the potential for generating “hallucinations,” and the internalization of data biases. If these obstacles can be overcome, the technology holds promise for breakthroughs in robotics and AI decision-making.
Meta’s series of actions—from restructuring teams and increasing investments to outlining a dual-model roadmap and exploring world models—demonstrate its determination to maintain a leading position in the next phase of artificial intelligence competition. However, the strategic balance between potentially shifting from open source to a proprietary model and the difficulty of overcoming frontier technologies like world models will continue to test this tech giant’s execution capabilities and technological foresight.
