Meituan has released a massive sparse Mixture of Experts model trained on domestic Chinese ASIC clusters to reduce reliance on Western hardware.
Beijing-based Meituan Inc. released its LongCat-2.0 large language model on June 30, 2026, featuring a 1.6-trillion-parameter architecture trained entirely on domestic Chinese compute clusters. As reported by SiliconANGLE, this release signals a significant pivot toward vertical integration within the Chinese AI sector.
The model employs a sparse Mixture of Experts (MoE) architecture similar to established frameworks like Mixtral and DeepSeek. It utilizes an internal router mechanism to activate a curated subset of expert parameters per token, which enhances inference efficiency by avoiding the activation of the entire model during every compute cycle.
With a 1.6-trillion-parameter scale, LongCat-2.0 provides a 1-million-token context window designed for high-density data processing. This capacity positions the model against other MoE systems that typically utilize 128,000-token context windows, prioritizing heavy-duty reasoning over smaller activation footprints.
Meituan integrated the model into its broader ecosystem to function as a core engine for AI agents and automated coding harnesses. The company specifically cited compatibility with tools like Claude Code, OpenClaw, and Hermes to facilitate repository-level edits and complex task orchestration.
The training methodology represents a strategic shift toward domestic AI Application-Specific Integrated Circuit (ASIC) clusters. By optimizing the model for local hardware, the company seeks to mitigate the operational risks associated with fluctuating access to high-end Nvidia Corp. graphics processing units.
Bernstein analysts noted in 2025 that Nvidia held approximately 40% of the Chinese AI chip market, though market share is projected to decline as domestic alternatives from firms like Huawei Technologies Co., Ltd. gain traction. Meituan’s reliance on domestic ASIC superpods suggests a move toward vertical integration within the Chinese hardware ecosystem.
The model’s deployment strategy requires high-density inference clusters rather than consumer-grade hardware. Due to the parameter count, the system is architected for data center environments where model parallelism can be managed across distributed compute nodes.
The technical architecture suggests that while the core reasoning logic remains portable, the performance optimizations are intrinsically tied to domestic silicon. This alignment provides a blueprint for enterprise-scale AI development that operates independently of traditional Western hardware supply chains.
Meituan’s transition into large-scale model development follows its 2023 acquisition of the startup Light Year Beyond for approximately $281 million. The release of LongCat-2.0 marks the culmination of internal development efforts that began in 2025 to build proprietary, high-capacity reasoning engines.
The significance of this release lies in the successful application of sparse MoE architectures to non-Nvidia hardware environments. By decoupling the model’s reasoning capabilities from specific CUDA-based dependencies, Meituan demonstrates that high-parameter models can achieve competitive performance on localized ASIC architectures.
This development underscores a broader trend among Chinese technology firms to insulate their research pipelines from international export controls. The ability to maintain a 1.6-trillion-parameter model on domestic infrastructure serves as a critical proof-of-concept for the scalability of Chinese semiconductor ecosystems in the face of global supply chain volatility.
Stakeholders in the AI research community are closely monitoring the performance benchmarks of this model against established Western counterparts. The shift toward domestic hardware optimization may force a reevaluation of how global AI development is measured, particularly regarding the necessity of specific high-end GPU architectures for training large-scale, sparse models.
Future performance benchmarks will determine whether the model maintains parity with closed-source systems like Google LLC’s Gemini or OpenAI’s GPT-5.5. Researchers will monitor how the model handles long-context tasks compared to existing industry standards in agentic workflows.
The reliance on domestic hardware will likely serve as a primary indicator for the viability of large-scale model training within China’s constrained semiconductor environment. Continued performance data will reveal the efficacy of these ASIC clusters in sustaining massive, sparse-model architectures over extended deployment cycles.
