Ant's Ling-3.2-flash Model Beats Claude Opus 5 on Key Benchmarks

Ant Group's Ling team has launched a new multimodal AI model called Ling-3.2-flash. It has 560B total parameters with 25B active ones, and supports a 1M token context window. It can take images, text, and video as input.

The model focuses on four skills: deep image and video understanding, long-running agent tasks, 3D scene understanding, and medical image analysis.

In official tests, it beat Claude Opus 5 (max) on three benchmarks, including OmniDocBench1.5 and Vision2Web. It also entered the top 10 on DesignArena with an Elo score of 1238.

The model uses a mixture-of-experts design, calling only part of its parameters to keep costs low. Its API costs $0.30 per million input tokens and $0.90 per million output tokens, the lowest rate in its class. It is live on Ant's Bailian platform and free for the first two weeks.