Meta signals Llama 4 arrival in 2026 with Liquid Transformers 2.0 architecture, aiming to reduce AI dependency on US cloud vendors

Meta is developing Llama 4 using a new Liquid Transformers 2.0 architecture (a way to structure AI models so they can adapt their processing based on the specific task they're solving) and plans to release it by 2026, positioning it as part of a "Sovereign AI" initiative to give countries and companies alternatives to relying on US-based AI infrastructure.

Liquid Transformers 2.0 allows the AI to dynamically adjust how much computing power it uses for different requests — simple questions use less resources, complex ones use more — which means lower operating costs and faster responses compared to fixed-size models that waste resources on easy tasks.

For business users and developers, this matters because a more efficient open-source alternative to closed models (like OpenAI's GPT or Google's Gemini) would lower costs to build AI features into their own products, and reduce reliance on paying per-API-call to US companies; for governments and enterprises in Europe and Asia, it offers a path to run advanced AI domestically rather than sending data overseas.