Meta prepares Avocado and Mango: powerful AI models will also be released in trimmed-down open versions

Meta is preparing two new advanced AI models for release in 2026: the large language model Avocado and the multimedia generator Mango. At the same time, the company plans to release more limited open-source versions of these systems, but without precise timelines and without promising full parity with closed releases.

Two models on the way

According to published information, Meta has two products of different classes in development. Avocado is described as a large language model, that is, a basic tool for text tasks: from dialogue and summarization to programming and document analysis. Mango is already a multimedia file generator.

The current description does not reveal whether it will primarily be about images, video, audio, or a combined system, but the vector itself is clear: the company is betting not only on text, but also on a broader set of formats. It is also important that both models belong to advanced systems and will remain closed for now. For the market, this is a signal that Meta is not planning to completely abandon its open-source strategy, but does not want to release the strongest versions into the public domain immediately after launch.

This approach is increasingly chosen by major players: publicly supporting ecosystem openness, but keeping the best capabilities in the commercial domain. This allows simultaneously collecting revenue from corporate scenarios and regulating access to the most expensive computational resources.

Open, but with limitations

Separate interest is drawn to the promise to release open-source versions on the same technological basis. This means that developers, research teams, and startups will likely be able to work not with completely separate models, but with derivative variants of Avocado and Mango. At the same time, Meta is signaling in advance: the functionality of such releases will be limited, which means the open layer of the ecosystem will not become an exact copy of flagship closed systems.

From a practical standpoint, this is a compromise between the scale of distribution and control over the most sensitive or expensive capabilities. The company can open the architectural foundation and a significant portion of quality, but at the same time retain for closed versions the most powerful modes, better multimodality, fine security tools, or commercially important optimizations. This format suits both the community and Meta itself: the former get material for development, the latter get space for monetization.

• Meta plans to release closed versions of Avocado and Mango in 2026
• Open releases will likely be built on the same foundation
• Their capabilities will be limited compared to flagship versions
• The company has not yet named exact release dates for open variants

Why Meta needs this

The news directly fits into the company's broader strategy: to distribute its AI models as widely as possible throughout the world. For Meta, this is not simply a matter of reputation in the open-source community. The more teams that build products on its technological stack, the stronger the ecosystem, the higher the tool compatibility, and the more noticeable the company's influence on the developer market.

At the same time, the scheme no longer looks purely ideological: the company clearly separates showcase closed models and a more accessible external layer for mass adoption. For Meta, this is especially important in competition for developers. If the company can offer strong open releases, even if not top-tier in functionality, it will remain a notable provider of basic infrastructure for applications, local deployments, and custom solutions.

And closed Avocado and Mango will serve as a demonstration of the upper tier of its AI platform—the level for which global business markets are usually willing to pay separately.

What this means

Meta is essentially betting on a two-tier market model: the best AI systems are under the company's control, more accessible versions are for mass distribution and ecosystem. If the scheme with Avocado and Mango works, users will get more open tools, and the market will have another example of how open source and closed frontier models now coexist. It seems that exactly such a balance is becoming the new standard for large AI companies in the coming years.