The Zhitong Finance App learned that on November 6, the Ant Bering Model announced the official launch of LING-3.2-Flash. The model has already been launched in Lingguang 2.0, which focuses on reliable execution and delivery in real production environments.
The total parameters of LING-3.2-Flash are 560b, activate 25B in a single step, use a 7:1 hybrid attention architecture, natively combine visual perception and spatial understanding, and are equipped with 1M original growth context. The core highlight is the visual feedback self-test+abnormal adaptive recovery mechanism: the model can use the screen and interactive interface as observation input, compare differences in structure, lighting, and layout, and automatically correct output results; when problems such as errors or tool call failures are encountered, it can locate problems and retry compensation to ensure continuous progress of long-term tasks.
At the evaluation level, the model scored 1,238 points in DesignArena's image-to-frontend front-end code generation evaluation. It can achieve the ability to convert visual design drafts to front-end pages, two-dimensional drawings to 3D models, manipulate professional software such as Blender, interpret papers, and generate runnable code. Under Hopper architecture GPU and about 30K long context conditions, the average decoding speed for a single user can reach 280 tokens/s, and a single token takes about 3.54 ms, taking into account task quality and inference efficiency.
Application scenarios cover multi-agent collaboration, design and production, R&D engineering, document processing, etc., and support the division of labor and collaboration between multiple heterogeneous agents to complete complex delivery. At the same time, the official frankness model still has shortcomings in long-term planning, 3D asset details, etc., and will continue to iterate and optimize and open source related results in the future. Currently, users can experience the model through Ling Studio and Ant Group's digital technology model service platforms, and model weights will also be released soon.