Original title: meituan-longcat/LongCat-2.5
A native multimodal MoE model open-sourced by the Meituan LongCat team, officially reported to have 1.6 trillion total parameters and approximately 480 billion active parameters per token. Building on LongCat-2.0, it combines visual understanding with agentic programming, targeting complex codebase development, full-project scaffolding, creative design, and 3D tasks. Images are converted into discrete tokens at their native resolution and aligned with text in the same semantic space; training employed long-horizon agentic reinforcement learning and multi-teacher distillation. Expert weights and activations underwent W4A4 quantization, enabling native INT4 computation on domestic AI accelerator cards. It supports both thinking and non-thinking modes.
Why it matters Among domestic open-source large models, few integrate vision and agentic programming; weights are MIT-licensed and deployable on domestic accelerator cards
Rising score 57 · Likes 57 · Downloads 0
Reasons
- Official evaluations cover programming agents, general agents, and vision, with scores approaching mainstream closed-source models in several metrics; scores without asterisks are self-tested, and the official statement notes they are for reference only
- The demo illustrates a multi-turn workflow involving image recognition followed by tool invocation
- Weights are MIT-licensed, and direct conversation is available on the official website longcat.ai
Usage
- Online trial: longcat.ai; for NPU deployment, see the longcat-2.5-npu branch documentation in SGLang-FluentLLM on GitHub
- The official recommendation is to enable enable_thinking and preserve historical reasoning content (save_history_reasoning_content)
Risks
- The model is extremely large and cannot run on personal devices; the model card provides only NPU deployment instructions
- Most evaluations are official self-tests
- This entry has not been run; refer to the model card for performance details