One-sentence summary
Comparing Meituan's LongCat-2.5 (1.6T total, ~48B active, 1M context, trained on domestic compute) with DeepSeek and Qwen on parameter architecture, context and compute route.
The Models
On November 2, 2026 Meituan launched LongCat-2.5, joining China's trillion-parameter LLM lineup alongside DeepSeek and Qwen.
Core Comparison
| Model | Total params | Active params | Context | Compute route |
|---|---|---|---|---|
| Meituan LongCat-2.5 | 1.6T | ~48B/token | 1M tokens | Fully domestic compute training |
| DeepSeek family | Trillion-class MoE | Sparse activation | Long context | Multiple clouds, domestic chips |
| Qwen family | Multi-size MoE | Sparse activation | Long context | Multi-cloud, multi-chip |
Key Differences
LongCat-2.5 stresses fully domestic-compute training and native multimodal long-horizon agents; DeepSeek and Qwen have mature open-weight and multi-size product lines with broader developer ecosystems.
Sources
- Meituan: Meituan launches LongCat-2.5
- Huxiu: Meituan launches 1.6T LongCat-2.5