Meta Plans Open-Source Versions of New AI Models Under Wang Leadership

Quick Facts

• Meta developing two proprietary frontier models codenamed Avocado (LLM) and Mango (multimedia generator) expected to launch this year
• Llama 4 Maverick will be an open-source model with 400 billion parameters, delivering inference costs as low as 10% of GPT-4o
• Meta invested $14.3 billion in Scale AI acquisition, bringing founder Alexandr Wang to lead superintelligence efforts

Meta is preparing to release open-source versions of new AI models developed under Alexandr Wang’s leadership, according to recent reports. The company plans to eventually offer these models through open-source licenses while maintaining proprietary versions.

The tech giant is developing two frontier models set to launch this year. The first is an LLM codenamed Avocado. The second is a multimedia file generator known internally as Mango.

Wang joined Meta after the company purchased 49% of Scale AI in June 2025. He stepped down as CEO of Scale but remained on the board. Meta announced it will invest $14.3 billion in Scale AI, overseen by Meta Superintelligence Labs.

The move represents a strategic shift toward a hybrid approach. Wang indicated that some of Meta’s largest new models will remain proprietary. This marks a departure from Meta’s earlier full openness strategy.

“AI is one of the most revolutionary technologies of our time, with unlimited possibility and far-reaching influence on how people, businesses and governments succeed,” Wang said.

Meta’s upcoming Llama 4 Maverick will be an open-source large language model with 400 billion parameters. Benchmarks show it competing directly with GPT-4 while being significantly more efficient to run.

The company said the first family of models is designed to help it catch up to rivals after its last Llama 4 family fell significantly behind. Meta believes it will have areas of strength that appeal to consumers despite acknowledging its new models may not be competitive across all dimensions.

Before releasing open-source versions, Meta wants to ensure they don’t add new safety risks. The open-source models may lack some neural networks that power proprietary editions or have scaled-down parameter counts.

Meta’s $14.8 billion investment in Scale AI represents nearly 25% of its projected $65 billion AI spending for 2025. Analysts view the deal as a massive acquisition designed to bypass internal development bottlenecks.

The strategy allows Meta to maintain its position as a force for democratizing AI access while competing with companies like OpenAI and Anthropic that focus on government and enterprise customers.

Read more: Report: Meta developing open-source versions of upcoming AI models