Meituan Releases Large-Scale Open Model "LongCat-2.0"

Chinese delivery giant Meituan has officially released the large-scale language model "LongCat-2.0" on GitHub, Hugging Face, and its own platform. The model is provided under the MIT License, allowing free use including commercial applications. With 1.6 trillion parameters, it adopts an architecture called "MoE (Mixture-of-Experts)," which combines multiple specialized models.

Interestingly, this model had already made its presence known in the developer community before its official release. Over the past two months, it was registered on OpenRouter under the anonymous name "Owl Alpha" and consistently ranked high on the global developer rankings. The official release has now revealed its true identity as Meituan's LongCat-2.0. OpenRouter is an API platform that allows unified access to multiple AI models and serves as a venue where developers worldwide evaluate models based on actual performance.

Technically, it is characterized by a maximum context window of 1 million tokens (the length of text that can be processed at once). The pricing structure is API-based pay-per-use, with standard prices set at $0.75 per 1 million input tokens and $2.95 per 1 million output tokens. Additionally, promotional discounted pricing of $0.30 for input and $1.20 for output is being applied, which is lower compared to major models from Google, OpenAI, and Anthropic. Furthermore, it features a mechanism where API calls for cached context are processed free of charge.

One key reason this model attracts particular attention is its training environment. LongCat-2.0 was trained on a cluster composed exclusively of over 50,000 domestically produced Chinese ASIC (application-specific integrated circuit) chips—semiconductor chips specialized for particular processing tasks. Amid continued US export restrictions limiting the supply of high-performance chips like NVIDIA GPUs to China, completing large-scale model training using only domestic chips holds significant industrial importance.

In AI model development, the performance of chips used in training directly affects model quality and training costs. Consequently, most cutting-edge models assume US-made high-performance GPUs as their foundation. This case challenges that assumption and serves as an indicator of China's efforts toward semiconductor self-sufficiency in the AI industry.

Competition in open-source models has intensified since 2024, with major Chinese-origin models being released successively. LongCat-2.0 combines the accessible MIT License conditions with a unique free cached context pricing strategy, positioning itself as an option for enterprises building their own AI systems. Given that Meituan operates diverse services spanning delivery, e-commerce, and maps, one can discern a strategy to strengthen its own AI infrastructure while simultaneously offering it externally.

Future points of interest include whether the capability that maintained top rankings for two months under anonymity will withstand broader developer validation following official release. Additionally, whether models trained exclusively with domestic chips can further improve performance will serve as a crucial indicator for gauging the overall direction of China's AI development going forward.