Ant Ling today launched Ling-3.2-flash, a native multimodal model with 560B total parameters but only 25B activated at a time. It comes with a 1M context window and follows the MoE approach—during forward passes, only part of the parameters are used, preserving knowledge capacity while reducing per-step cost. This "large skeleton, small activation" design is the key to balancing capability.
Four capabilities arrive together The model architecture natively connects vision and language, handling images, text, and video alike. The official release highlights four strengthened areas: complex visual understanding for deep reasoning over images and video; multimodal long-horizon Agent that stays focused on goals across multi-step tasks and makes and executes decisions based on visual and textual input; 3D understanding to parse spatial relationships; and medical multimodal capabilities covering image interpretation and report generation.
Three benchmarks beat Opus 5 Ant's evaluation shows that Ling-3.2-flash outperforms Claude Opus 5 (max) on three benchmarks: OmniDocBench1.5 scores 92.21 (vs. 90.56), ZEROBench(Sub) 43.71 (vs. 39.52), and Vision2Web 85.86 (vs. 80.20). In DesignArena's image-to-frontend arena, it entered the top ten with an Elo of 1238, sharing the stage with Meta, DeepSeek, and Anthropic.
Cheapest in its class, free for the first two weeks API pricing is $0.30 per million input tokens and $0.90 per million output tokens, rated by DesignArena as the lowest rate in its performance tier. The model is now available on the Ant Bailian platform, is compatible with OpenAI and Anthropic protocols, and is free for the first two weeks after launch.
