Ant Bailing released a native multi-modal model Ling-3.2-flash. Here are the specifications first: MoE architecture, total parameters 560B, only 25B activated each time, context 1M.
The total parameter is large and the activation is small. The advantage is that the knowledge capacity is retained and the overhead of single-step reasoning is reduced. Domestic models have basically followed this route in the past two years.
Officials say it focuses on strengthening four areas: complex image and video reasoning, multi-step multi-modal Agent tasks, 3D spatial understanding, and medical image interpretation and report generation.
The most conspicuous thing in the promotion is that the three benchmarks beat Claude Opus 5 (max level):
OmniDocBench1.5 is 92.21 vs. 90.56, ZEROBench(Sub) 43.71 vs. 39.52, Vision2Web 85.86 vs. 80.20. Received an Elo of 1238 in DesignArena’s image-to-front-end arena, ranking among the top ten.
Here you need to see clearly who your opponent is. Now Claude's flagship is Opus 5.5. Compared with the previous generation, winning only means that it has caught up to the position half a year ago. Moreover, all three items were tested by Ant itself, and the selected areas were all its main focus areas such as document analysis and visual conversion to web pages.
What really matters is the price. The API input is US$0.30 per million tokens and the output is US$0.90, which is rated as the cheapest in its category by DesignArena. The first two weeks of launch are free, and it is compatible with OpenAI and Anthropic protocols. Changing the interface is basically changing the address.
For a team that does document OCR and convert screenshots to code, it’s worth running a round of real business in the past two weeks. If you don’t run for free, it will be in vain.