美团开源万亿参数大模型 LongCat-2.0，基于五万卡国产算力集群训练
英文摘要
On June 30, Meituan released and open-sourced LongCat-2.0, a next-generation trillion-parameter large language model. The model has 1.6 trillion total parameters, with an average activation of 48 billion (dynamic range 33B–56B), and was fully trained and inferred on a domestic computing cluster of 50,000 cards. Pre-training data exceeds 30 trillion tokens, covering Chinese, English, multilingual and code, and the model natively supports a 1-million-token context length. This release demonstrates large-scale model training on domestic hardware and makes a massive open-source model available to the research community.
中文摘要
6月30日，美团发布并开源了新一代万亿参数大模型 LongCat-2.0。该模型总参数1.6万亿，平均激活约480亿（动态范围330亿~560亿），在五万卡国产算力集群上完成全流程训练与推理。预训练数据规模超过30万亿 tokens，覆盖中英文、多语言和代码，原生支持100万超长上下文。此举展示了在国产硬件上进行大规模模型训练的能力，并向研究社区开放了庞大的开源模型。
关键要点
LongCat-2.0 has 1.6 trillion total parameters and uses a Mixture-of-Experts design with an average activation of 48B (dynamic range 33B–56B).
总参数1.6万亿，采用MoE架构，平均激活约480亿参数（动态范围330亿~560亿）。
The model was fully trained and can be inferred on a cluster of 50,000 domestic computing cards.
模型在五万张国产计算卡的集群上完成全流程训练与推理。
It was pre-trained on over 30 trillion tokens covering Chinese, English, multilingual, and code data, with native 1M context length.
预训练数据超30万亿tokens，覆盖中英、多语言及代码，原生支持100万上下文长度。
Meituan has open-sourced LongCat-2.0, providing the research community with a large-scale model.
美团将LongCat-2.0对外开源，为研究社区提供了大规模的模型。