🤖 AI Curated
Meituan has released LongCat-2.0, a 1.6 trillion parameter MoE coding model, under an MIT license. It completed training and inference using a cluster of approximately 50,000 Chinese-made ASICs, without Nvidia GPUs. This near-frontier-level model had already dominated the top ranks on OpenRouter under the disguised codename 'Owl Alpha'.
Meituan, China's largest delivery and lifestyle services company, has unveiled 'LongCat-2.0,' a 1.6 trillion (1.6T) parameter open-source coding-specialized model, through its AI research organization 'LongCat'. It's not just a large model; the core of the buzz is 'where and how it was trained.' This model completed everything from pre-training to inference services, end-to-end, in a cluster composed of approximately 50,000 Chinese-made AI accelerators (domestic ASICs), without using any high-performance GPUs from Nvidia. The industry is hailing it as 'the first instance of a trillion-parameter model implemented end-to-end solely with Chinese chips.'
Structurally, it employs an MoE (Mixture of Experts) approach, dynamically activating only about 33B to 56B (averaging around 48B) parameters per token, rather than using all 1.6 trillion parameters at once. This ensures inference efficiency despite its massive scale. Furthermore, it incorporates its self-developed 'LongCat Sparse Attention' and a 'zero-computation expert' technique that skips calculations for simple tokens, allowing it to natively process ultra-long contexts of 1 million (1M) tokens. This design targets agent-based software tasks that require reading and inferring from entire large-scale codebases.
Its performance is also presented as 'near-frontier' level. It scored 59.5 on SWE-bench Pro, a practical coding benchmark; 70.8 on Terminal-Bench, a terminal task benchmark; and 77.3 on the multilingual SWE-bench. Some reports indicate that its SWE-bench Pro score slightly surpasses GPT-5.5 (58.6). The pre-training data is reportedly between 30 and 35 trillion tokens, encompassing multilingual data and code, including Korean.
The model's potential was already proven before its official release. A preview version of LongCat-2.0, quietly uploaded to the open-source API routing platform OpenRouter under the disguised codename 'Owl Alpha,' climbed to the top ranks globally based on actual usage calls. Statistics showed it processed approximately 10 trillion tokens per month (an average of about 559 billion tokens per day), a 242% surge compared to the previous month. In other words, developers were already using it extensively, judging solely by its performance, without even knowing its brand.
The license is the relatively open MIT, and the API is provided via OpenAI and Anthropic-compatible endpoints (promotional pricing around $0.30/input million tokens and $1.20/output million tokens). However, the exact timing of the full release of its weights varies across different media outlets. Some reported that they are already available from Hugging Face's 'meituan-longcat' organization, while other reports stated 'weights coming soon,' meaning reconfirmation of actual download availability is needed from official channels.
In summary, LongCat-2.0 presents two narratives simultaneously. One is a geopolitical signal: 'China can create frontier-level large models using only domestic chips, despite U.S. semiconductor export restrictions.' The other is a practical signal: 'Another low-cost, open-source powerhouse has emerged in the coding agent market.'
For domestic development teams and startups, LongCat-2.0 is proof that 'frontier-level coding models can emerge even without Nvidia,' and it means there's one more low-cost alternative to the Claude and GPT series. Its MIT license, OpenAI and Anthropic-compatible API, and 1 million token context make it attractive for agent-based automation that can ingest entire in-house codebases. However, due to the nature of its training data and Chinese infrastructure, verification is needed before adoption in security, data governance, and regulation-sensitive domains. Crucially, whether the weights are actually released will determine its value as an open-source project. It's also noteworthy that the competition for AI semiconductor self-sufficiency is an issue that will soon impact the domestic cloud and chip ecosystem.
🤖 AI-curated from multiple sources. Verify accuracy with the originals (sources).