Meta readies next‑generation “Mango” and “Avocado” AI models for 2026 launch
- Meta is developing an image and video generation model codenamed **Mango** and a new text large language model called **Avocado** for a targeted first-half 2026 release.[1][3]
- Both models are being built inside **Meta Superintelligence Labs**, led by Scale AI co-founder Alexandr Wang, and were outlined in an internal Q&A with staff.[1][3][5]
- Mango is designed to advance multimodal image and video generation, while Avocado is aimed at significantly improving coding and reasoning capabilities over Meta’s current LLMs.[1][2][8]
- The projects are described internally as Meta’s first flagship models following a major AI reorganization and are intended to compete with frontier systems from OpenAI and Google.[1][4][7]
- Meta has discussed Mango and Avocado in internal roadmaps and blog updates but has not yet issued a formal public product launch announcement or fixed release date for the models.[1][6][8]
Meta is preparing two next-generation artificial intelligence models, an image and video system known internally as “Mango” and a text-based large language model dubbed “Avocado,” with an internal roadmap targeting a launch in the first half of 2026, according to company documents and media reports.
New models Mango and Avocado targeted for 2026
Reports from The Wall Street Journal, Bloomberg and other outlets say Meta is building Mango as a new image and video generation model and Avocado as a next-generation text large language model intended to improve on the company’s current Llama-based systems.[1][3][4] An internal Q&A held with employees and later described in media coverage outlined a roadmap to release the models in the first half of 2026.[1][3][7] According to these accounts, the timeline is an internal target rather than a publicly announced launch date and could still shift as development progresses.[1][4]
Meta has not issued a standalone public product announcement for Mango or Avocado but has referenced the projects in an update on its Meta Superintelligence Labs initiative and broader AI roadmap.[8] The company has confirmed its focus on more capable multimodal models and stronger coding performance but has not disclosed detailed technical specifications, model sizes or training data sources for Mango and Avocado.[1][8]
Role of Meta Superintelligence Labs and leadership
Both Mango and Avocado are being developed within Meta Superintelligence Labs, an internal organization created as part of Meta’s AI restructuring and focused on building more advanced foundation models.[1][4][8] The lab is led by Alexandr Wang, co-founder of data-labeling firm Scale AI, who joined Meta to oversee its push into what the company describes as superintelligence research and large-scale model training.[1][5][8]
During the internal Q&A cited in multiple reports, Wang and Meta chief product officer Chris Cox presented Mango and Avocado as early flagship projects for the lab and central to Meta’s next phase of AI development.[1][3][5] The efforts follow a period of turnover and reorganization in Meta’s AI ranks, including the departure of chief AI scientist Yann LeCun to start a separate company and the exit of several researchers who had been recruited into the new lab.[1][4]
Technical goals: multimodal generation and coding
According to descriptions shared with employees and reported externally, Mango is intended to strengthen Meta’s capabilities in multimodal generation, particularly for images and video that can respond to text prompts and potentially other forms of input.[1][2][3] The model is part of what Meta leadership has described as an exploration of "world models" that can better understand and reason about visual information, plan sequences of actions and operate in more complex environments.[1][2]
Avocado is framed internally as a successor to Meta’s current text LLMs with a specific focus on software development and reasoning tasks.[1][2][5] Wang told employees that Meta aims to make the model "much better at coding" and capable of handling more complex instructions and tools than its predecessors.[1][2] While Meta has not disclosed benchmark results, the company’s internal positioning suggests it views Avocado as a direct response to coding-focused models from OpenAI, Google and other competitors.[3][4][7]
Position in Meta’s AI strategy and competitive landscape
Media reports describe Mango and Avocado as the first major models to emerge from Meta’s revamped AI organization and as a key test of whether the company can keep pace with rivals in frontier AI research.[1][4][7] Meta’s existing Meta AI assistant has been widely distributed through its social apps, but analysts note it has not yet been viewed as clearly ahead of competing assistants in quality.[1][4] The company’s vast social platforms, however, give it access to distribution channels and user feedback that could be used to refine future models.[1][6]
The new models are intended to compete with leading systems such as OpenAI’s GPT series and Google’s Gemini models by offering more capable multimodal generation and improved coding features, according to coverage in the Financial Times, Reuters and other outlets.[4][6][7] These reports say Meta leadership has framed Mango and Avocado as part of a broader strategy to close perceived gaps with top competitors while continuing to release some models under open or partially open licenses, though the licensing approach for these specific systems has not yet been clarified.[4][6][8]
Companies mentioned
Meta Platforms Inc., which operated as Facebook, Inc. until its October 2021 rebranding, is a technology enterprise focused on developing innovative products that empower people globally to connect and share with their …