LongCat-2.0 ran as ‘Owl Alpha’ on OpenRouter before reveal
Meituan on June 30 unveiled LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts model that ran anonymously for two months on OpenRouter as ‘Owl Alpha’ and led Hermes Agent by call volume.
Meituan revealed LongCat-2.0 on June 30. The company identified the model as the system that operated on OpenRouter under the name ‘Owl Alpha’ for about two months and ranked first by monthly call volume on the Hermes Agent workspace, second on Claude Code and third on OpenClaw deployments.
LongCat-2.0 is a sparse mixture-of-experts model with 1.6 trillion parameters. Meituan reports the model activates roughly 48 billion parameters per token on average, with activation ranging from about 33 billion to 56 billion depending on query complexity. The model supports a 1 million-token context window.
The architecture combines a sparse attention method called LongCat Sparse Attention, designed to handle very long conversations by focusing computation on the most relevant parts of context, and an N-gram embedding approach that lets the model represent common multiword phrases without significantly increasing model size. After training, Meituan integrated three specialist subsystems-Agent for tool use, Reasoning for problem solving and Interaction for dialogue-and uses a routing layer to select which subsystem mix to apply to each request.
Meituan reported that pretraining ran across more than 50,000 domestically produced accelerators and consumed over 35 trillion tokens. The company described the training and deployment as completed end-to-end on domestic ASICs, and said the run finished without rollbacks or irrecoverable loss spikes.
On benchmarks, LongCat-2.0 scored 59.5 on SWE-bench Pro, compared with 58.6 for GPT-5.5 and 54.2 for Gemini 3.1 Pro. On FORTE, it scored 73.2, tied with Claude Opus 4.6 and below GPT-5.5’s 77.8. In brief coding tests, including a game-building task, the model produced usable outputs but trailed some higher-ranked models on quality; low API cost made iterative refinement more affordable in those tests.
Meituan listed standard API pricing at $0.75 per million input tokens and $2.95 per million output tokens, with a launch promotion cutting rates to $0.30 per million input and $1.20 per million output. Cached context reads are offered at no charge. The company also offers bulk token packs of 1 billion tokens for about $60.
Developers can access LongCat-2.0 through Meituan’s OpenAI- and Anthropic-compatible API endpoints and through agent platforms that have integrated the model. Meituan has not released model weights for self-hosting; GitHub and Hugging Face repositories linked to the project state ‘model weights coming soon’ without a firm delivery date.
The material on GNcrypto is intended solely for informational use and must not be regarded as financial advice. We make every effort to keep the content accurate and current, but we cannot warrant its precision, completeness, or reliability. GNcrypto does not take responsibility for any mistakes, omissions, or financial losses resulting from reliance on this information. Any actions you take based on this content are done at your own risk. Always conduct independent research and seek guidance from a qualified specialist. For further details, please review our Terms, Privacy Policy and Disclaimers.
Articles by this author
