Meituan Deploys LongCat 2.5 Using Sparse Mixture-of-Experts Architecture

Meituan released LongCat 2.5 on Monday, a 1.6-trillion-parameter multimodal model engineered for long-horizon autonomous coding tasks, marking the second consecutive iteration trained entirely on domestic Chinese AI accelerators. According to a report by TechTimes, this release provides evidence that export controls on high-end Western chips have not prevented domestic labs from achieving significant model scale, though the architectural methodology behind this performance remains the primary focus for researchers.

The system utilizes a Mixture-of-Experts architecture that activates approximately 48 billion parameters per input token, effectively decoupling total model capacity from inference compute requirements. This sparse activation strategy allows the model to function efficiently on hardware clusters estimated to be Huawei Ascend 910C units, which offer roughly 60% of the per-chip performance of Nvidia H100 accelerators. By routing tokens through specialized sub-networks rather than the full model, Meituan minimizes the compute gap that would otherwise render a dense model of this scale unviable on non-Nvidia hardware.

Meituan implemented a proprietary mechanism termed LongCat Sparse Attention to mitigate output discontinuity and quadratic scoring bottlenecks observed in previous architectures. The framework integrates Streaming-aware Indexing for memory-efficient token selection alongside Cross-Layer Indexing and Hierarchical Indexing to optimize recall across deep network layers. These components allow the model to maintain coherence during complex, multi-step coding sequences that require sustained attention over long context windows.

The model incorporates 135 billion N-gram Embedding parameters, a sparse block designed to prevent performance degradation that typically limits standard sparse activation methods. Training spanned over 35 trillion tokens and reportedly achieved stability without irrecoverable loss spikes, despite the memory constraints inherent in the domestic hardware stack. This engineering achievement suggests that the model’s stability is a direct result of the sparse routing logic rather than raw compute throughput.

LongCat 2.5 is designed for autonomous agent workflows, enabling the system to plan, execute code, and perform iterative error recovery without human intervention. The model retains a 1-million-token context window, though the default serving configuration is capped at 256K to balance latency and memory overhead. This design choice reflects a pragmatic approach to serving large-scale models where memory bandwidth remains a bottleneck for real-time inference.

Meituan has integrated the model into its Miaoshow merchant workbench to facilitate operational tasks for hotel and restaurant operators. The training pipeline includes a failure-recycling loop that converts real-world production errors into training data, allowing the model to adapt to specific edge cases encountered in commercial deployments. This methodology ensures that the model evolves based on actual user interactions rather than relying solely on static, curated evaluation datasets.

Meituan claims performance parity between Claude Opus 4.8 and Claude Opus 5, though these results remain unverified by independent third-party benchmarks. Previous iterations of the LongCat family have shown significant discrepancies between in-house SWE-bench Pro scores and independent evaluations, complicating direct comparisons with Western frontier models. The lack of shared benchmarks between this model family and global competitors makes it difficult to ascertain the true reasoning capabilities of the system.

The absence of public service-level agreements or transparent data retention policies remains a critical hurdle for enterprise adoption. While the model demonstrates technical efficacy in coding benchmarks, the lack of third-party security audits leaves questions regarding data governance and server-side processing locations unanswered for international users. Researchers note that the legal framework governing data handling in China imposes structural requirements that may conflict with the privacy expectations of global enterprise clients.

Future adoption will likely depend on the release of independent benchmark data and the clarification of enterprise data processing agreements. The broader research community will continue to monitor whether the LongCat 2.5 architecture can maintain its performance edge as the model scales further on constrained domestic hardware. Until independent evaluators gain access to the model, the performance claims should be viewed as directional indicators rather than settled scientific results.