Meta has also upgraded the vision encoder in Llama 4. While it’s based on MetaCLIP, the company trained it separately alongside a frozen Llama model, helping it better align with the language ...