Meituan has open sourced its 1.6T-parameter LongCat-2.0 AI model under the MIT License, giving developers and enterprises a commercially friendly agentic coding model with a 1-million-token context window.
Meituan has open sourced LongCat-2.0, releasing the 1.6-trillion-parameter Mixture-of-Experts (MoE) model on GitHub, Hugging Face and its own platform under the permissive MIT License. Previously known as the anonymous Owl Alpha model, LongCat-2.0 topped OpenRouter developer rankings before its identity was revealed.
The enterprise-friendly MIT License allows developers to modify, redistribute and integrate the model into proprietary software without being required to open source derivative works, making it one of the year’s most commercially flexible open-source AI releases.
Designed for autonomous coding, software engineering, AI agents and repository management, LongCat-2.0 supports a native 1-million-token context window. It was trained entirely on a cluster of more than 50,000 Chinese ASIC chips, demonstrating that near-frontier AI models can be developed without relying on Nvidia GPUs.
Before its public unveiling, Owl Alpha processed around 10.1 trillion monthly tokens, averaging 559 billion tokens daily, while recording 242% month-on-month growth. It ranked among OpenRouter’s global top three models and secured the top spot on Hermes Agent Workspace.
LongCat-2.0 scored 59.5 on SWE-bench Pro, narrowly outperforming GPT-5.5’s 58.6, underscoring its strength in software engineering. The model also introduces LongCat Sparse Attention, Zero-Compute Experts and the MOPD post-training framework to improve long-context efficiency and agent performance.
Meituan says organisations can self-host the model to reduce API dependence, improve data privacy and lower operational costs through free context-cache hits, positioning the company as an emerging global open-source AI infrastructure provider.
