Can domestic AI chips support trillion-parameter large models? The ceiling has officially been shattered. Meituan has unveiled LongCat-2.0, a self-developed Mixture-of-Experts (MoE architecture) model. Boasting a staggering 1.6-trillion total parameters and activating approximately 48B parameters per token, it natively supports a 1M long-context window. Most remarkably, the entire pipeline from training to inference was completed with zero Nvidia dependency, marking the world's first trillion-parameter model to achieve a full-loop closed-loop on domestic Chinese silicon.
Before its official announcement, the model was quietly battle-tested in the wild. Under the alias Owl Alpha, it was deployed on OpenRouter, where it rapidly dominated developer charts. It claimed the No. 1, No. 2, and No. 3 spots in monthly call volumes on Hermes, Claude Code, and OpenClaw respectively, proving to be the absolute favorite among global Agent developers.
In empirical tests, #LongCat-2.0 demonstrated outstanding long-context processing and logical reasoning. When evaluated on a mixed-language corpus of tens of thousands of words comprising complex financial and scientific papers, the model successfully executed pinpoint information retrieval within 1-second retrieval limits, demonstrating robust comprehension.
In software engineering tasks, LongCat-2.0 operates with top-tier Agent-native coordination. Given a 13k-star pure HTML/CSS/JS GitHub repository of the game 2048, the model autonomously generated a 7-step plan and executed a comprehensive cyberpunk visual redesign, board expansion to 5x5, and a step counter in just 12 minutes. It even seamlessly completed a full React migration of the game, showcasing clean and modular rewriting capabilities.
Furthermore, the model stands out in cost efficiency and autonomous search. Under the same prompt for physical simulations compared with top frontier models, LongCat-2.0 utilized merely 9,004 tokens. This efficiency is highly optimized by Meituan's Cache-hit pricing strategy, which substantially slashes overhead for developer iterations.
Accomplishing this on a cluster of 50,000 domestic chips represents a monumental system engineering feat. Due to smaller VRAM limitations and the lack of high-speed inter-node bandwidth compared to NVLink, training a 1.6-trillion parameter #MoE model requires highly intricate parallelization. The team overcomed this by redesigning core operators like FlashAttention and building customized communication topologies, optimizing throughput and proving that non-Nvidia hardware is fully viable for world-class AI workloads.
[AgentUpdate Depth Analysis] The emergence of Meituan's LongCat-2.0 signals a pivotal shift in the AI Agent ecosystem, proving that trillion-parameter MoE models can be efficiently trained and deployed on non-Nvidia domestic hardware. For advanced AI Agents, which require frequent tool calls, continuous code-generation, and multi-step reasoning, inference cost and context length are critical bottlenecks. LongCat-2.0 addresses these pain points by offering native 1M context support and highly optimized token pricing (enhanced by cache-hit discounts). By demonstrating robust performance under the "Owl Alpha" alias on platforms like #OpenRouter, it has proven itself capable of powering demanding developer-centric agent environments (such as Claude Code). This development suggests that the future of AI Agent deployment is moving toward decentralized, heterogeneous hardware architectures, democratizing access to trillion-parameter capabilities while significantly lowering operational costs for complex multi-agent workflows worldwide.
