Ant Bailing has released its large-scale model, Ling-3.1-flash, as open-source software. This model is tailored for handling complex tasks and executing long-range operations, leveraging a hybrid linear attention architecture. Boasting a total of 560 billion parameters, the model activates only approximately 25 billion parameters per token, supporting a context window of up to 1 million tokens. Through architectural and quantization enhancements, the model substantially cuts down on GPU memory consumption while maintaining a balance between performance, speed, and cost, showcasing remarkable intelligence efficiency.

Practical tests demonstrate that Ling-3.1-flash is capable of independently accomplishing tasks such as screening AI industry news to create topic selection dashboards, autonomously developing complete city-building games over extended sequences, and generating interactive teaching webpages for Fermat's principle. These achievements highlight its significant practical value in real-world applications.

Ant Bailing does not prioritize benchmark scores; instead, it emphasizes applications in real-world scenarios. Looking ahead, the company will persist in refining the model's long-range execution capabilities and intelligence efficiency. It will integrate feedback from industry-specific scenarios, such as healthcare and finance, to bolster its professional capabilities and promote large models as dependable tools for industrial productivity.