With 8 trillion parameters and 1 million word contexts, Kimi K3 has raised the ceiling of open source large models to the highest in the world. Before the World Artificial Intelligence Conference has even begun, domestic large models have released a blockbuster.

2. With 8 trillion parameters and 1 million word contexts, Kimi K3 has raised the ceiling of open source large models to the highest in the world. Before the World Artificial Intelligence Conference has even begun, domestic large models have released a blockbuster. As the 2026 World Artificial Intelligence Conference and the High-Level Conference on Artificial Intelligence Global Governance are about to be held, Dark Side of the Moon officially released the new generation model Kimi K3 on the 16th, with a parameter scale of 2.8 trillion, making it the world's largest open source model with parameters. This is not only a numerical summit, but also means that the open source camp has left the closed source giants behind in terms of size for the first time. Kimi K3 does not just rely on stacking parameters to support the scene. According to Beijing Dark Side of the Moon Technology Co., Ltd., this model natively supports visual understanding and has an ultra-long context window of 1 million words, which is equivalent to swallowing the entire monograph in one breath and still remembering what was written at the beginning. It has been specifically optimized for use in complex task scenarios such as software engineering, knowledge work, in-depth research, and multi-modal understanding. The processing capabilities of complex tasks have thus reached a new level - from reading pictures to recognizing objects to long-range code engineering. K3 attempts to break down the hard bones that used to be broken into multiple steps and squeeze them into the same model. The capital side has already taken notice. CITIC Construction Investment gave a set of confident judgments in the research report: the volume of global large model calls is expected to increase month-on-month for 11 consecutive weeks, domestic models have steadily occupied the forefront of traffic due to their high cost performance, and the revenue expectations of leading overseas manufacturers have also been raised accordingly. What is more worth pondering is the ebb and flow of the pace of commercialization - domestic manufacturers are obviously speeding up, DeepSeek is preparing to launch time-of-use pricing, and Doubao has been launched for multiple membership levels; on the other hand, Anthropic has tightened the access restrictions of Claude. Under the intertwining of compliance and geographical factors, the logic of domestic substitution has been further illuminated. Along this main line, the Cailianshe theme database has sorted out several listed companies that resonate with it. In the field of AI, 360 takes 360 Intelligent Brain and 360 Security Large Model as its core layout. The self-developed 360 Intelligent Brain, a universal large-scale model with hundreds of billions of parameters, already has multi-modal understanding, logical reasoning, code generation and other capabilities. Kunlun Wanwei announced on July 2 that its subsidiary Tiangong AI achieved a historic breakthrough in the second quarter of 2026. The annual recurring revenue (ARR) of its AI Native model and product business has crossed the US$800 million mark. When the open source behemoth with 2.8 trillion parameters comes into being, the prosperity of domestic large models is being pushed to a new high point visible to the naked eye. via AI News (author: AI Base)