On the same day that Huang Renxun supported open source, Ant Bailing threw out the 124B "execution model"

Intelligent Things Author | Wang Han Editor | Mo Ying In the past week, two major events in the Silicon Valley AI circle have pushed the battle over the AI ​​route to a climax. First, Jen-Hsun Huang’s first While Silicon Valley is still arguing about "should it be open source" and "should it hit the brakes?", in the Chinese AI circle across the ocean, open source and advancement have never seemed to be a multiple-choice question. From DeepSeek to Zhipu, from Ali Tongyi to Ant Bailing, the open source pace of China's large models has never stopped. July 24th - the same day that Huang Renxun tweeted in support of open source, Ant Bailing officially released the new generation of native hybrid inference model Ling-3.0-flash. The model is available for free API calls on OpenRouter and Bailing official platforms simultaneously. The free period ends at 23:00 on August 3. The model weights will be officially open source after the free period ends. In the Token call volume ranking, Ling-3.0-flash jumped from 9th to 6th place within 5 days since its release. The call volume on July 29 was only 0.5% lower than DeepSeek V4 Pro. The total parameter amount of Ling-3.0-flash is 124B, and the amount of activated parameters is only 5.1B. Compared with its flagship thinking model Ring-2.6-1T, which was open sourced in May, the total parameter amount is about 12.4% of the latter, and the amount of activated parameters is only 8.1% of the latter. However, 11 of the 12 benchmark tests performed better than Ring-2.6-1T. Silicon Valley giants have been debating the question "Can open source be both secure and powerful?" Ant Bailing gave its own answer with its products. 1. Use small to win big, 11 out of 12 tests surpassed the trillion flagship. Let’s take a look at the results of Ling-3.0-flash first: In the benchmark test, Ling-3