Another heavyweight player has joined the competition in the field of large AI models. Recently, Meituan officially announced its latest developed trillion-parameter large model - LongCat-2.0, and declared it to be fully open-sourced, aiming to promote industry technology exchange and application development.

LongCat-2.0 demonstrates the strong potential of domestic computing clusters. It is reported that this model was trained and inferred on a five-card domestic computing cluster, not only verifying the reliability of domestic software and hardware in large-scale distributed computing, but also providing a reference pattern for building ultra-large models in the industry. From a technical specification perspective, LongCat-2.0 has a total parameter count of 1.6T, using a dynamic range design (activation parameters about 48B, dynamic range covering 33B to 56B). This design ensures a vast knowledge reserve while maintaining the flexibility and efficiency of inference.

In terms of data support and context processing capabilities, LongCat-2.0 also performs well. Its pre-training data scale exceeds 30T tokens, deeply covering Chinese and English corpora, and integrating multilingual and high-quality code data, achieving cross-domain logical understanding. More impressive is that the model supports 1M ultra-long context natively, which means it can maintain high coherence and accuracy when handling tasks such as long document analysis and complex codebase construction.

By choosing to open-source LongCat-2.0, Meituan has undoubtedly injected new vitality into the developer community. As an open-source model with a trillion-parameter scale, it not only marks Meituan's continuous efforts in the research and development of AI infrastructure, but also contributes an essential computing foundation to the prosperity of the domestic large model ecosystem. In the future, as this model is widely applied, we can expect more innovative AI applications based on ultra-long context understanding to emerge.