Meituan releases LongCat-2.0, a trillion-parameter AI model, now open-source

Meituan just released its new AI model LongCat-2.0, which has over a trillion parameters and is now open-source. The model uses 1.6 trillion parameters in total, with about 48 billion active at any time. It supports super long text up to 1 million tokens and was trained from scratch on a cluster of 50,000 Chinese-made GPUs. A preview version is already available on OpenRouter and longcat.ai, and it's now one of the top three most-used models on OpenRouter. LongCat-2.0 was trained on more than 30 trillion tokens of data, including Chinese, English, other languages, and code. The team cut monthly failure rates by over 70% and boosted training efficiency by 1.5 times, reaching a steady daily output of over 1 trillion tokens. The model uses a smart attention method called LongCat Sparse Attention to handle long texts efficiently, and it can dynamically assign computing power to complex tokens while skipping simple ones. Its MOPD design combines three expert groups for tasks like tool use, math reasoning, and conversation. In coding tests, LongCat-2.0 scored 59.5 on SWE-bench Pro, beating Gemini 3.1 Pro (54.2), GPT-5.5 (58.6), and Claude Opus 4.6 (57.3). It also got 77.3 on SWE-bench Multilingual, close to Claude Opus 4.6 (77.8), and 70.8 on Terminal-Bench 2.1. In real-world office tasks, it scored 78.8 on RWSearch, 73.2 on FORTE, and 79.9 on BrowseComp, matching top closed-source models.