Ant Group Releases Ling 3.2 Flash Multimodal Model with 1M Context
Ant Group has released its native multimodal model, Ling-3.2-Flash, which features 560 billion total parameters, a 1-million-token context window, and competitive performance across several prominent vision and language benchmarks. The model challenges established industry competitors by outperforming Anthropic's Claude Opus 5 on select vision-related benchmarks while offering very competitive API pricing and a limited-time free trial. Ling-3.2-Flash utilizes a mixture-of-experts architecture with approximately 25 billion activated parameters per run to balance capacity and compute costs, while supporting API integration compatible with OpenAI and Anthropic protocols.
## BACKGROUND
Mixture of Experts (MoE) is a machine learning architecture where only a subset of network parameters is activated for each given input, enabling large model capacities with lower computational overhead. Benchmarks like OmniDocBench 1.5 and Vision2Web evaluate specific multimodal capabilities such as complex document parsing and visual website development.