According to CoinMeta, Ant Bird's released ling-3.2-flash model brings native multimodal capabilities to the 560B flagship model, supporting image, text, and video inputs, with a focus on enhancing complex visual understanding, 3D interaction, and multimodal agent abilities. The new model continues to utilize the previous generation's ling-3.1-flash's total of 560 billion parameters, 25 billion activation parameters, and 1 million token contexts, expanding into native visual understanding. In official demonstrations, the new model can create web pages based on images, operate Blender to generate 3D scenes, and can also read papers containing formulas and charts PDF, converting the algorithms within into executable code. Ant's self-tests show that the new model outperforms the previous model by more than 5max in three evaluations: document understanding, complex visual reasoning, and creating web pages from images. Specifically, it scored 85.86 in vision2web, compared to the previous model's 80.20 in Opus. On third-party DesignArena's image-generation front-end rankings, ling-3.2-flash ranked 10th with 1238 elo. Currently, the model is available for web page experience and API; it is free for the first two weeks.