meituan-longcat/longcat-2.0

LongCat-2.0 is a very large AI language model developed by Meituan, a Chinese technology company. Like GPT or Claude, it can read and write text, answer questions, write code, and follow complex instructions. It has 1.6 trillion total parameters, a measure of model size, with about 48 billion active at any given moment. It uses an architecture called Mixture of Experts, where only a portion of the model activates for each request. What sets this model apart is its focus on extremely long context windows. It was trained on sequences up to one million tokens long, which means it can read and reason over an entire large codebase or a very long document in a single pass. Most publicly available models top out at 200,000 tokens or less. The model was built specifically for coding and agent tasks. It is integrated with tools like Claude Code and other AI coding assistants, meaning it can help with reading whole repositories, making large-scale code edits, and running automated tasks across many steps. Benchmark scores show it competing closely with leading models from Google and Anthropic on coding-agent evaluations. The training used custom AI chips called ASIC superpods rather than standard graphics cards. The training run covered more than 35 trillion tokens across millions of accelerator-days without any training failures or rollbacks. Model weights are available on HuggingFace and ModelScope, a Chinese model hosting platform. The license is MIT, meaning you can use it freely including for commercial purposes. The README links to a technical blog post with full architecture details.