Meituan LongCat open-sources LongCat-2.5-W4A4, a 1.6T-parameter multimodal agent model

LongCat-2.5-W4A4 is a native multimodal mixture-of-experts model for long-horizon agentic tasks, combining visual understanding with agentic coding for complex codebase development, full-project builds, creative design, and 3D tasks. It builds on LongCat-2.0 with long-sequence training, large-scale reinforcement learning, multi-teacher learning, and infrastructure optimizations on domestic AI accelerators. The model card reports benchmark scores across coding, general agent, and visual tasks, with some scores cited from public reports and others measured in-house.

The model uses native-resolution visual features converted into discrete tokens and jointly aligned with text in a shared semantic space. Training combines real-world engineering tasks and reproducible agent failures with efficient rollout strategies and safeguards against reward hacking. Multi-teacher learning uses full-vocabulary KL distillation and Tree Attention. Supernode-aware parallelism, efficient sharding, and communication–computation overlap achieve 86% higher per-device throughput than DDP baseline in reported long-sequence SFT tasks. KV cache optimization and LSA nearly double decode KV capacity without increasing TPOT on the tested workload.