China's first domestically produced card training model with trillions of dollars, open source!

Zhidongxi Author | Eggplant Editor | Cheng Qian Zhidongzhi reported on July 6 that Meituan today announced the open source of LongCat-2.0, a large model with trillions of parameters, and simultaneously opened the inference code for deep optimization of domestic computing chips. The total parameters of this model reach 1.6 trillion, with an average activation of about 48 billion parameters. It is the first trillion-parameter model in the industry to complete full-process training and inference on a 50,000-ka domestic computing cluster. ▲LongCat-2.0 is open source (picture source GitHub) From the official data, LongCat-2.0 has strong comprehensive capabilities. In SWE-bench Pro, which examines deep engineering capabilities, it scored 59.5, ahead of Gemini 3.1 Pro (54.2), GPT-5.5 (58.6) and Claude Opus 4.6 (57.3); in SWE-bench Multilingual, which examines programming languages, it scored 77.3, ahead of Claude Opus 4.6 (77.8) is flat; it scored 70.8 in the real terminal command interaction evaluation Terminal-Bench 2.1. ▲Score comparison between LongCat-2.0 and various models (picture source: LongCat) In terms of real scene tasks, LongCat-2.0 scored 78.8 in the search agent evaluation RWSearch, 73.2 in the productivity scene evaluation FORTE, and 79.9 in the BrowseComp score, all reaching or approaching the level of cutting-edge closed-source models. According to the actual measurement of Zhixixi, LongCat-2.0 can better understand the user's meaning in terms of long text output, and can generate readable long text according to user requirements. In terms of programming, the code generation of this model is faster, but sometimes it is unstable and there will be problems with the picture. Moreover, this model also has good creative ability and 3D animation scene generation ability, and can understand the user's meaning and respond immediately. In addition, LongCat-2.0 also has strong logical reasoning capabilities, and the steps are clear and concise when reasoning about questions. LongC