Ant Group launches Ling-3.0-flash model, small size but big performance

Ant Group's Bailing AI released the new Ling-3.0-flash model today. It has 124B total parameters but only 5.1B active ones, much smaller than the previous flagship model. Yet it matches or beats the old one in reasoning, following instructions, and handling long texts. The model uses a mixed linear attention design from the start, blending KDA and MLA layers. With just 5.1B active parameters, it boosts efficiency in long-context tasks and complex AI agent scenarios. For real-world use, Ling-3.0-flash trained on over 10,000 interactive environments, handling coding, general tasks, and deep research. It also uses SGLang HiCache and Mooncake caching to cut response time by 60-80% for long inputs. Right now, you can try the model for free on OpenRouter and Bailing's platform until August 3rd. After that, the model weights will be open-sourced.