LongCat-2.5
Multimodal MoE model for long-horizon agentic tasks
LongCat-2.5 is a native multimodal mixture of experts model built by Meituan LongCat. It features 1.6 trillion total parameters with 48 billion activated per token. The system combines visual understanding with agentic coding to handle complex development and design tasks.
- 1.6 trillion total parameters with 48 billion active
- Native multimodal support for text and vision
- W4A4 quantization for efficient INT4 computation
- Specialized in long-horizon agentic coding tasks
- Deployable on domestic NPU platforms
What it is
LongCat-2.5 is a large language model designed for extended agentic workflows. It operates as a native multimodal system, processing both text and visual inputs within a shared semantic space. The model contains 1.6 trillion parameters in total, though it activates only about 48 billion for each token processed. This architecture allows it to manage complex tasks such as full-project code builds, creative design work, and three-dimensional modelling. It serves as an upgrade to the earlier LongCat-2.0 version, focusing on reliability in long-horizon execution.
How it came about
The model builds directly upon the capabilities of LongCat-2.0. Its development involved long-sequence training and large-scale reinforcement learning. The team employed multi-teacher learning strategies, where specialized agents guided the model through distillation processes. Infrastructure optimizations were made specifically for domestic AI accelerators. These steps aimed to improve the efficiency of training and the reliability of end-to-end task execution in real-world engineering scenarios.
Inside the model
The architecture uses a mixture of experts design. Visual features are converted into discrete tokens at native resolution and aligned with text. Training utilized supernode-aware parallelism and efficient sharding to boost throughput. The model applies W4A4 quantization to expert weights and activations, enabling INT4 computation. This approach reduces data movement and communication overhead. Fused kernels and optimized weight layouts further enhance serving efficiency. The system supports both thinking and non-thinking modes, allowing users to balance performance with token efficiency.
What it can do
The model performs across coding, general agent, and visual benchmarks. In Terminal Bench v2.1, it scored 88.4, while achieving 38.5 in Terminal Bench v4.0. It recorded 64.6 on DeepSWE v1.1 and 82.1 on PaperBench. For visual tasks, it reached 96.5 on MathVision with tools and 85.2 on VideoMME. The system handles codebase question answering, refactoring, and test writing. It also supports 3D code generation and spreadsheet analysis. Scores indicate strong performance in long-context reasoning and multi-turn instruction following.
How to use it
Weights are available for download via Hugging Face. The model can be deployed on NPU platforms using SGLang-FluentLLM. It supports standard API interactions with options to enable or disable thinking mode. Users can configure parameters such as temperature and top_p. The quantization strategy allows for efficient deployment on domestic hardware without significant accuracy loss compared to unquantized versions.
Safety and testing
Training included safeguards against reward hacking during reinforcement learning. Multi-teacher learning incorporated interaction teachers to strengthen safety protocols. The model uses bitwise-identical token log probabilities across training and inference to maintain alignment. Red-teaming and evaluation harnesses were used to test reliability in automated tasks.
Questions people ask about LongCat-2.5
What is the parameter size of LongCat-2.5?
The model has 1.6 trillion total parameters. However, it uses a mixture of experts architecture, so only about 48 billion parameters are activated for each token processed.
Can LongCat-2.5 process images?
Yes, it is a native multimodal model. It converts visual features into discrete tokens and aligns them with text in a shared semantic space, allowing it to understand images and perform visual tasks.
What hardware is required to run LongCat-2.5?
The model is optimized for domestic AI accelerators and can be deployed on NPU platforms. It uses W4A4 quantization to enable efficient INT4 computation, reducing memory and communication requirements.
Good reading and sources
There is no English Wikipedia article on LongCat-2.5 yet, so the facts on this page come from the sources listed above: huggingface.co/meituan-longcat/LongCat-2.5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.