🐱 LongCat 2.0: Meituan's 1.6T MoE Stealth Model Revealed
The mysterious 'owl-alpha' on OpenRouter was Meituan's LongCat-2.0 — a 1.6T parameter Mixture-of-Experts model using Huawei Ascend 910C superpods and LongCat Sparse Attention.
📋 Article Summary
- 🔑 What Is LongCat-2.0 — 1.6T total params, ~48B activated per token, MoE architecture
- 🚀 Hardware — Trained entirely on Huawei Ascend 910C ASIC superpods
- ⚡ Attention — LongCat Sparse Attention (LSA), evolution of DeepSeek Sparse Attention
- ⚠️ Sparsity Trade-off — 97% sparsity means scaling experts yields diminishing returns
- 💡 Stealth Deployment — Previously available as 'owl-alpha' on OpenRouter before removal
Model Specifications
LongCat-2.0 is a large-scale Mixture-of-Experts language model with 1.6 trillion total parameters and approximately 48 billion activated parameters per token. The model uses LongCat Sparse Attention (LSA), an evolution of DeepSeek's sparse attention mechanism.
Technical Architecture
Training Infrastructure
Both the full training run and large-scale deployment are built entirely on AI ASIC superpods — specifically Huawei Ascend 910C processors.
LongCat Sparse Attention (LSA)
Builds on LongCat-Flash architecture, pushing further on parameter efficiency for long-context training and inference.
Multi-Token Prediction (MTP)
Extended three strategies to the 3-step MTP module for accelerating speculative decoding.
Sparsity Analysis
The MoE sparsity has crossed the sweet spot. With ~97% sparsity even without N-gram Embedding, scaling up experts by 135B parameters yields negligible performance gains. This suggests the efficiency frontier for current MoE designs.
Hardware Specs
| Component | Spec | Details |
|---|---|---|
| Processor | Huawei Ascend 910C | 48 machines × 8 processors, 64GB HBM each |
| Network | 200 Gbps RDMA | Inter-processor communication |
| Total Parameters | 1.6T | ~48B activated per token |
Key takeaway: Previously stealth-deployed as 'owl-alpha' on OpenRouter, LongCat-2.0 represents a major Chinese MoE effort trained on domestic ASICs. Available at longcat.chat with weights pending HuggingFace release.