Meituan has officially released LongCat-2.5, the latest generation of its native multimodal large model built for long-horizon agent tasks. The model uses a unified token representation for vision and text, activates roughly 48 billion parameters per token out of 1.6 trillion total, and is tuned for domestic Chinese AI accelerators. In head-to-head benchmarks it beats Claude Opus 5 on AutomationBench and ranks first on BrowseComp.
Architecture and Scale
LongCat-2.5 is a major iteration on LongCat-2.0. It represents visual signals as multi-level discrete tokens so the backbone processes vision and text in one unified token form, deepening the fusion of "native visual understanding" and "agentic programming."
- Total parameters: 1.6 trillion.
- Active parameters per token: about 48 billion.
- Long-sequence SFT single-card throughput improved 86% versus baseline through a reworked parallelism and memory layout for accelerators with under 80GB of memory.
Benchmark Performance
Against overseas frontier models including Claude Opus 5, GPT-5.6-Sol, GPT-5.5, and Claude Opus 4.8, LongCat-2.5 leads several sub-leaderboards:
- AutomationBench: 51.8, surpassing Claude Opus 5 (50.3).
- BrowseComp: 91.9, ranking first.
- 3DCodeBench (text-to-3D): 55.4, a narrow lead over peers.
- Terminal-Bench v4.0: 38.5.
Domestic Hardware Optimization
The system targets China's domestic compute environment, breaking through long-context training and W4A4 inference bottlenecks. Fine-tuned training-inference numeric consistency and long-context rollout optimization support large-scale long-horizon reinforcement learning at high efficiency.
Key Data Points
- 1.6 trillion total parameters
- about 48 billion active parameters per token
- AutomationBench 51.8 (vs Claude Opus 5: 50.3)
- BrowseComp 91.9 (rank 1)
- 3DCodeBench 55.4
- Terminal-Bench v4.0 38.5
- plus 86% long-sequence SFT single-card throughput versus baseline
Sources: Zhidx, AIBase, AIComing, threerouter
