Apple has unveiled details of a new model called Manzano dedicated to understanding and generating images at the same time, a complex field that has long been an obstacle for open-source models compared to advanced commercial systems such as those developed by Google and OpenAI.

The model, which means apple tree in Spanish, has not yet been put into public use, and the company has not provided a demo of it, but has only published a paper of low-resolution image samples that have been used to compare its performance with open-source models such as Deepseek Janus Pro and commercial systems such as GPT-4o and Gemini 2.5 Flash Image Generation, also known as Nano Banana.

Apple emphasizes that open-source models often have to choose between the power of image analysis or the quality of their generation, while commercial models combine the two tasks together.

This is due to the nature of image processing, as understanding requires continuous data streams, while generation requires dividing images into discrete symbols, which usually leads to inconsistencies within the linguistic model.

To solve this dilemma, Apple has created a hybrid design based on a common image codec that produces two types of symbols, continuous codes that are used to understand images, and separate codes that are dedicated to generating them. This approach allows for less conflict between the two tasks, and gives the model greater flexibility in handling complex texts and graphics.