On July 3, Meituan officially released its next-generation foundation model LongCat-2.0, featuring an MoE architecture with 1.6 trillion total parameters and approximately 48 billion activated parameters per token.
"Zero NVIDIA Content": A Milestone for Domestic Compute
The model's most distinctive feature is that the entire pipeline from training to inference was completed on domestic compute clusters, with zero NVIDIA content. This represents not only a true full-stack breakthrough for domestic chips + domestic models but also places it directly in the global first tier.
Before official release, LongCat-2.0 was anonymously listed on the OpenRouter platform, with total calls ranking in the top three; in Claude Code Agent scenarios, its call volume ranked second globally, trailing only Claude Opus 4.8.
Objective Assessment: Not the Absolute Best, But Breakthrough in Significance
Technical community evaluations place its core Agent capabilities close to Claude Opus 4.6, lagging behind the latest Opus 4.8; Coding ability slightly exceeds Zhipu GLM-5.1 but falls behind GLM-5.2 released in June.
However, when adding the qualifier "zero NVIDIA content," these relative disadvantages become qualitative breakthroughs. While domestic large models have previously featured "domestic" narratives — some using domestic chips for inference, others for post-training — full training-plus-inference pipeline at trillion-parameter scale is a first.
Editor's Note: LongCat-2.0's significance lies not in "being the strongest," but in "achieving global first-tier performance without any NVIDIA chips." This proves that China's AI compute "backup plan" can not only function but also deliver competitive results. Combined with DeepSeek-V4's successful adaptation to Huawei Ascend and repeated market enthusiasm for domestic chip stocks, a dual-wheel-drive pattern of "algorithm autonomy + compute autonomy" is taking shape.