Meituan is pushing LongCat-2.0 to full open source under the MIT license.

Meituan open-sources its LongCat-2.0 MoE model under the MIT license, featuring 1.6T parameters, a 1M context window, and native integration with Claude Code.

After an initial release in late June, Meituan is fully opening the floodgates for LongCat-2.0: model weights and inference code are now public, under an unrestricted MIT license. The model is a Mixture-of-Experts with a total of 1.6 trillion parameters, approximately 48 billion of which are activated per token, with a context window of one million tokens.

Training and deployment rely entirely on AI-dedicated ASIC superpods rather than the Nvidia ecosystem, a choice the company presents as a demonstration of frontier-scale training on alternative hardware. Architecturally, LongCat-2.0 introduces LongCat Sparse Attention, an evolution of sparse attention designed to accelerate the processing of long contexts, and an N-gram Embedding module that expands the representation space to better leverage each parameter.

The model is designed for agentic uses and integrates directly with harnesses like Claude Code, OpenClaw, and Hermes, with scores that Meituan claims are on par with leading proprietary models for code, search, and task execution. Deployment is validated on both GPUs and NPUs. Weights and code are accessible on HuggingFace, GitHub, and ModelScope, with a testing interface at longcat.ai.