Meituan open-sources LongCat-2.0, a 1.6T-parameter LLM trained on Chinese chips
On 30 June 2026, Meituan released and open-sourced LongCat-2.0, a 1.6 trillion-parameter large language model that the company says is the industry's first trillion-parameter model to complete full training and inference entirely on a domestic Chinese compute cluster.
Architecture. LongCat-2.0 is a Mixture-of-Experts model with 1.6T total parameters, dynamically activating roughly 33B–56B (averaging ~48B) per token. It was pretrained from scratch on more than 30 trillion tokens and natively supports a 1 million-token context window.
Domestic hardware. The model was trained on a cluster of around 50,000 domestic compute cards, with no Nvidia or AMD GPUs in the pipeline. Meituan did not name its chip supplier but said it used the Huawei Collective Communication Library to improve training stability; analysts point to Huawei Ascend or Cambricon as the likely hardware.
Context. The release lands amid US export controls on advanced AI chips, and is positioned as evidence that frontier-scale training is now feasible on non-Nvidia hardware in China. The model is geared towards agentic coding, scoring 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, with weights published on Hugging Face.
Read more: SCMP · LongCat AI
