Meituan Open-Sources LongCat-2.0: 1.6T Params, Zero Nvidia, All Domestic Chips

Meituan trained a trillion-parameter model on 50,000 domestic chips, then open-sourced it.

On July 7, Meituan officially open-sourced LongCat-2.0, its new foundation model. With 1.6 trillion total parameters in MoE architecture and ~480B activated parameters per token, the model's most distinctive feature is that the entire pipeline — from training to inference — runs on a 50,000-card domestic compute cluster with zero Nvidia content.

Key Facts

• Total parameters: 1.6 trillion (1.6T)
• Activated parameters: ~480B per token
• Architecture: MoE (Mixture of Experts)
• Context window: Natively supports 1 million tokens
• Compute: 50,000-card domestic chip cluster, zero Nvidia throughout
• Open-sourced: Model weights + inference code optimized for domestic chips

This is the industry's first trillion-parameter model fully supported by domestic chips from training through inference. Previously, domestic chips were primarily used on the inference side, with training still heavily dependent on Nvidia GPUs. LongCat-2.0 demonstrates that the domestic chip + domestic model combination can now compete head-to-head with top international standards.

Why It Matters

Against the backdrop of geopolitical restrictions on Nvidia GPU exports to China, domestic compute training capability has been an industry focus. The prevailing view was that domestic chips lagged behind Nvidia in cluster interconnect, software ecosystem, and training stability, making trillion-parameter model training impractical.

LongCat-2.0's release breaks this perception. Training a trillion-parameter model on a 50,000-card cluster requires engineering capability across communication bandwidth, fault recovery, checkpoint management, and data parallelism. Meituan's simultaneous release of inference code optimized for domestic chips means other developers can deploy the model directly on domestic hardware.

Open Source Significance

Meituan's full open-source release (weights + optimization code) continues the trend established by DeepSeek, GLM, and other open-source models. The expanding open-source ecosystem gives developers more self-hosting options, reducing dependence on closed APIs.

Notably, LongCat-2.0's native 1-million-token context support provides significant advantages in document analysis, code understanding, and long conversation scenarios. Combined with the cost advantages of domestic compute, LongCat-2.0 may form differentiated competitiveness in enterprise use cases requiring long-context processing.

---

本文由无人日报（Deskless Daily）编译员整理发布。