LongCat-Flash
Efficient 560B parameter MoE model for agentic tasks
Meituan LongCat Team released LongCat-Flash in September 2025. This 560 billion parameter mixture of experts model activates roughly 27 billion parameters per token. It targets high efficiency and strong performance in agent-based workflows.
- 560 billion total parameters with 27 billion active per token
- Trained on 20 trillion tokens in just 30 days
- Achieves over 100 tokens per second inference speed
- Strong performance in agentic tool use benchmarks
- Open-sourced weights for public research
What it is
LongCat-Flash is a large language model built by the Meituan LongCat Team. It contains 560 billion parameters but uses a mixture of experts design to keep computation low. For each token processed, it activates between 18.6 and 31.3 billion parameters, averaging about 27 billion. The model was released in September 2025 with a focus on agentic tasks, which involve using tools and reasoning through complex steps. It supports a context window of 128,000 tokens and aims for high speed during inference.
How it came about
The Meituan LongCat Team developed the model to address scaling challenges in large AI systems. They submitted a technical report in early September 2025 detailing their approach. The team trained the model on more than 20 trillion tokens over a period of 30 days. Their goal was to create a foundation model that balances computational cost with advanced capabilities in reasoning and code generation. The checkpoint was made available to the public to support further research.
Inside the model
The architecture uses zero-computation experts to allocate budget dynamically based on token importance. A PID-controller adjusts expert bias to maintain a steady load. Shortcut-connected MoE blocks help overlap computation and communication, reducing bottlenecks. Training relied on hyperparameter transfer from smaller models and a model-growth initialization method. The team employed deterministic computation to ensure reproducibility and detect data corruption. Post-training involved synthetic data generation for complex agentic scenarios.
What it can do
Benchmarks show competitive results against other leading models. On MMLU, it scored 89.71, and on MMLU-Pro, it reached 82.68. Mathematical reasoning scores include 96.40 on MATH500 and 70.42 on AIME24. The model performs strongly in agentic tool use, scoring 73.68 on the telecom subset of the Tau-squared Bench. It also handles coding tasks well, with an 88.41 pass rate on Humaneval+. Inference speeds exceed 100 tokens per second.
How to use it
The model weights are open-sourced for community use. Developers can run it using adapted versions of SGLang and vLLM. The team reports an inference cost of $0.70 per million output tokens. Chat templates are provided for single and multi-turn interactions, including specific formats for tool calls. This allows integration into applications requiring function calling and structured outputs.
Questions people ask about LongCat-Flash
How many parameters does LongCat-Flash have?
The model has 560 billion total parameters. However, it uses a mixture of experts architecture, so only 18.6 to 31.3 billion parameters are active for any given token, averaging around 27 billion.
What is the context length of the model?
LongCat-Flash supports a context window of 128,000 tokens. This allows it to process and reason over long documents or extended conversations without losing track of earlier information.
Is LongCat-Flash free to use?
Yes, the model checkpoints are open-sourced. You can download the weights and run them locally or on your own infrastructure. The developers also provide integration guides for popular inference engines.
Good reading and sources
There is no English Wikipedia article on LongCat-Flash yet, so the facts on this page come from the sources listed above: huggingface.co/meituan-longcat/LongCat-Flash-Chat, arxiv.org/abs/2509.01322. 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.