Trillion parameters become standard configuration, domestic AI large models launch "close combat" battle.

Two years ago, when the yet-to-be-launched DeepSeek stirred up a price war with the attitude of a "price butcher," perhaps few would have thought that domestic large-scale models could narrow the gap with foreign top models to within 3 months in just two years, and at the same time have the right to raise prices. Recently, Chinese large-scale models have made consecutive breakthroughs. On July 17, the dark side of the moon released the Kimi K3 model, which has 28 trillion parameters and a contextual window of 1 million words, making it the largest open-source weight model in the world. On July 19, Alibaba Qianwen launched the Qwen3.8-Max preview model, and announced that the Qwen3.8 will be released and open-sourced soon, with a parameter scale of 24 trillion, "possibly the most powerful model after Fable 5." With trillion-level parameters and performance comparable to top closed-source models, domestic AI now has the confidence to raise prices. The application programming interface calling fee for the K3 model has increased significantly compared to the previous K2.6 price, making it the most expensive domestic large model currently. In the eyes of industry insiders, this is an important signal indicating that domestic large-scale models are transforming from cost-effective substitutes to technological front-runners. However, on this crowded track, the distance between players is constantly shrinking, and can almost be described as "close-combat." When technological advantages can be quickly overturned in a matter of weeks or even days, how can large model companies prove their uniqueness? This not only determines the market valuation they will receive, but also determines who will have the last laugh in this "survival of the fittest" competition.