Meituan has announced LongCat 2.0, an advanced model based on the Mixture-of-Experts (MoE) architecture, boasting a total parameter count of 1.6 trillion. By utilizing LongCat Sparse Attention technology, the model is capable of efficiently processing ultra-long sequences with a context window of up to 1 million tokens.
What Happened
Meituan developers trained LongCat 2.0 on 35 trillion tokens using specialized AI-ASIC superclusters instead of traditional GPUs. The model operates on the MoE principle, where despite a total scale of 1.6 trillion parameters, approximately 48 billion parameters are actively engaged to process each individual token. The model demonstrated high efficiency in programming and agentic planning tasks, as evidenced by its results in the SWE-bench and Terminal-Bench benchmarks.
Context
Unlike most modern frontier-scale models that rely on NVIDIA infrastructure, LongCat 2.0 was created with an emphasis on hardware independence. The use of custom AI ASICs allows for the optimization of training and inference processes for large-scale models, offering an alternative scaling path outside the established GPU ecosystem.
Why It Matters for the Industry
For the industry, this release demonstrates the viability of alternative paths for training frontier-scale models on specialized hardware. This reduces critical dependency on NVIDIA and stimulates the development of architectures optimized for ultra-long context and autonomous agent tasks.
Why It Matters for Users
Users gain access to a powerful tool for coding and deep analysis of documentation or repositories via an API that is compatible with OpenAI and Anthropic formats. This allows for the rapid integration of advanced agentic pipelines into existing workflows without the need to own expensive proprietary infrastructure.
What Is Not Yet Known / Limitations
At this time, detailed data regarding latency, inference costs, and API security requirements for industrial use are unavailable. Additionally, the status of releasing model weights on Hugging Face for local deployment has not been specified.
Sources
Author
Look at AI, Editorial Team