LongCat-2.0

LongCat-2.0 is a 1.6-trillion-parameter mixture-of-experts foundational large model developed by Meituan, officially released on June 30, 2026, and positioned for long-context AI workloads. It is described as the first foundational large model in its class to be both trained and served entirely on domestic computing infrastructure, with both stages running on a cluster of 50,000 cards. That end-to-end domestic compute pipeline is the model's defining technical story, signaling that trillion-parameter scale training can be sustained outside of conventional hyperscaler hardware stacks.

Shortly after release, Meituan open-sourced LongCat-2.0 under the MIT license, publishing weights, an inference engine, and technical documentation on GitHub, Hugging Face, and ModelScope, which makes commercial reuse straightforward for smaller teams. The mixture-of-experts design combined with a very large context window points to practical fit for long-document reasoning, multi-turn agentic workflows, and retrieval-heavy tasks where large effective parameter counts matter more than dense inference cost. The domestic ecosystem responded in lockstep, suggesting the model is intended to serve as a backbone for organizations prioritizing sovereign compute supply chains.