Key facts
- Meituan officially launched LongCat-2.0, a 1.6-trillion-parameter AI model, on June 30.
- The model operated anonymously as Owl Alpha on OpenRouter, achieving top rankings in call volume.
- LongCat-2.0 utilizes a mixture-of-experts architecture with approximately 48 billion active parameters per token.
- It was trained end-to-end on domestic Chinese ASICs.
- API pricing is competitive at $0.75 per million input tokens and $2.95 per million output tokens.
Meituan has officially launched LongCat-2.0, a large-scale AI model with 1.6 trillion parameters. This model had been operating anonymously for two months on OpenRouter under the alias Owl Alpha, where it achieved top rankings in call volume across platforms like Hermes Agent, Claude Code, and OpenClaw.
LongCat-2.0 is a mixture-of-experts model that activates approximately 48 billion parameters per token, with this number fluctuating based on query complexity. Meituan emphasizes that this is the first trillion-parameter model trained and deployed entirely on domestic Chinese ASICs, highlighting a move towards reduced reliance on U.S. hardware for AI development. The company stated that the extensive pretraining run, involving over 35 trillion tokens on more than 50,000 accelerators, was completed without significant issues.
The model's pricing is a key selling point, with standard API access set at $0.75 per million input tokens and $2.95 per million output tokens, significantly undercutting competitors like GPT-5.5 and Claude Sonnet 5. Promotional rates are even lower. Meituan also offers a token plan for heavy users. In testing, LongCat-2.0 performed adequately for coding tasks, though it showed limitations in handling complex logic and scaling issues in game development scenarios.
Meituan attributes the model's efficiency to techniques such as LongCat Sparse Attention, which processes only relevant parts of long conversations, and an N-gram embedding system that enhances language understanding by recognizing phrases as single concepts. The model also integrates specialized systems for agentic coding, reasoning, and conversation, with a routing mechanism directing requests to the appropriate specialist. On benchmarks like SWE-bench Pro and FORTE, LongCat-2.0 achieved competitive scores, surpassing some established models in specific tasks.