The global AI arms race has entered a new phase of geopolitical defiance. Meituan, the Chinese services giant, has released LongCat-2.0, a frontier-scale language model that challenges the long-held assumption that trillion-parameter AI requires Nvidia's ecosystem to be viable. The most striking aspect of this release is the infrastructure: LongCat-2.0 was trained entirely on Chinese-made hardware, marking a significant milestone in operational independence.
MoE Architecture and Massive Scale
LongCat-2.0 utilizes a Mixture-of-Experts (MoE) architecture, boasting a total of 1.6 trillion parameters, with approximately 48 billion activated per token. This design allows the model to maintain high-level reasoning capabilities while optimizing computational costs during inference. The training process was an engineering feat, utilizing a cluster of over 50,000 domestic AI ASICs and processing more than 35 trillion tokens.The team behind LongCat had to overcome significant technical hurdles, including communication faults at scale, memory pressure, and numerical stability—challenges typically solved using Nvidia's proprietary software stack. By overcoming these obstacles, Meituan has proven that a fully domestic AI stack is technically viable for frontier-scale training.
Performance Analysis: Coding vs. General Reasoning
Positioned as an agentic coding model, LongCat-2.0 shows impressive results in specialized benchmarks. According to reports from The Decoder, the model outperforms Gemini 3.1 Pro and GPT-5.5 on SWE-bench Pro (59.5) and SWE-bench Multilingual (77.3), although it still trails behind Claude Opus 4.7 and 4.8.However, the gap is wider in general reasoning tasks. On benchmarks such as IFEval (90.0), IMO-AnswerBench (81.8), and GPQA-diamond (88.9), LongCat-2.0 falls short of the top Western flagship models. Despite this, achieving competitive coding performance on non-Nvidia hardware is a clear signal that the technical barrier to entry for trillion-parameter models has lowered.
