Meta plans to release open-source versions of its next-generation frontier AI models. The models are believed to be derived from two new proprietary models that Meta is developing internally under the codenames Avocado and Mango. Both proprietary versions are scheduled for release this year.
This is according to Axios, citing sources. According to Axios, Meta is working on two proprietary frontier models: Avocado, a large language model, and Mango, a multimedia file generator. The open-source variants are expected to be made available at a later date and are part of Meta’s effort to distribute its models as widely as possible worldwide.
Last December, Bloomberg reported that Meta would switch to a closed-source distribution approach for future LLMs. Now, however, the company appears to be sticking to its open-source tradition, albeit alongside proprietary versions. Llama 4, which Meta launched a year ago, is built on a mixture-of-experts architecture with 128 separate neural networks, each optimized for specific tasks.
Open-source, but not fully
The open-source versions are not expected to include all the features available in the closed editions. AI security plays a role here, suggesting that Avocado will be able to generate cybersecurity-related code. In that category, Meta is exercising extra caution with public releases.
Exactly what will be omitted remains unclear. Possible limitations include a reduced number of parameters, the absence of certain neural networks, or skipping post-training steps. The latter is not uncommon in open-source models: developers often release streamlined variants without full post-training.
It is also not yet clear whether the open-source versions will be released simultaneously with the closed variants or only afterward.
According to Axios, Meta does not expect its models to beat the competition on all fronts. Nevertheless, the models are said to have several “strengths.” One potential advantage: hardware efficiency. Many current frontier models are too resource-intensive for standard PCs. Meta could optimize its models for applications such as personal health and homework assistance.