Meituan has launched LongCat-2.0, an AI model with 1.6 trillion parameters, offering competitive pricing and innovative features to challenge existing models.
Key Highlights
- Meituan unveiled LongCat-2.0 on June 30, branding it the foundation of their "Owl Alpha" technology.
- The model ranked first on Hermes Agent and performed well against competitors based on call volume.
- API pricing significantly undercuts its rivals, starting at $0.75 per million input tokens.
On June 30, Meituan officially launched LongCat-2.0, a cutting-edge AI model boasting 1.6 trillion parameters. Previously anonymous, the model had been operating as "Owl Alpha" for two months on OpenRouter.
Utilizing a nimble training process, LongCat-2.0 employs approximately 48 billion parameters per token, dynamically adjusting this number based on the complexity of each task. This model has become notable for being the first of its kind trained solely on domestic Chinese ASICs, distancing itself from others that rely heavily on external technology.
One standout feature is the pricing structure. The standard API access comes at $0.75 per million input tokens and $2.95 for output tokens. Currently, promotional rates are even lower at $0.30 and $1.20, making it a compelling alternative to competitors such as GPT-5.5 and Claude Sonnet 5, whose prices are significantly higher.
In practical testing, LongCat-2.0 performed adequately in game development scenarios but did not reach the output quality of top competitors like Claude Fable. Its ability to generate engaging enemy waves was impressive; however, it struggled with gameplay fluidity when faced with numerous adversaries.
LongCat-2.0 showcases several advanced features, including an innovative attention mechanism to enhance processing speed and a new N-gram embedding system that improves the model's understanding of phrases. These advancements not only enhance linguistic capabilities but also maintain a compact model size.
The model integrates specialized systems for various tasks, optimizing performance by routing requests to the appropriate capabilities. In benchmark tests, LongCat-2.0 secured a respectable score of 59.5 on SWE-bench Pro, exceeding GPT-5.5 but falling short of the latest Claude models.
Meituan announced: "Introducing LongCat-2.0 š±, a model with 1.6T parameters, ready for agentic coding from the ground up!"
ā Meituan LongCat (@Meituan_LongCat) on June 30, 2026.
LongCat-2.0 is now accessible through Meituan's API and supports platforms like Hermes and Claude Code. However, self-hosting options are not yet available, with Meituan yet to release the model weights publicly.
