Ant group's ling 3.2 flash model launches, outperforming claude opus 5 on key benchmarks

Ant Group has unveiled its native multimodal model, Ling-3.2-flash, which features a total of 560B parameters and 25B activation parameters, boasting a 1M context window. The model natively integrates visual and language capabilities, supporting inputs from images, text, and video, with specific enhancements in complex visual understanding, multimodal long-range Agent functionality, 3D comprehension, and medical multimodal analysis. In evaluations released by Ant Group, Ling-3.2-flash surpassed Anthropic's Claude Opus 5 on three benchmarks. Specifically, in the DesignArena image-to-frontend arena, the model achieved an Elo score of 1238, placing it among competitors like Meta, DeepSeek, and Anthropic. The model's architecture balances large knowledge storage with controlled inference costs, typical of a Mixture-of-Experts (MoE) setup. Pricing for the API is set at $0.30 per million input tokens and $0.90 per million output tokens, which DesignArena noted is the lowest rate among models of comparable performance. The API is now available on the Ant Bailian platform, supporting OpenAI and Anthropic protocols, and is offered with a two-week free trial.