What is LongCat-Flash-Thinking
LongCat-Flash-Thinking is a large reasoning model with 560 billion parameters launched by the Meituan LongCat team. Based on a Mixture of Experts (MoE) architecture, it supports dynamically activating 18.6 billion to 31.3 billion parameters according to demand, balancing computational efficiency and performance. Through two-stage training consisting of long chain-of-thought cold-start training and large-scale reinforcement learning, the model possesses strong capabilities in formal reasoning, agentic reasoning, and more. It performs excellently in complex reasoning tasks such as mathematics, logic, and programming, and can provide efficient intelligent support for fields including scientific research, software development, and enterprise decision-making, promoting the inclusive development of AI technology.
The LongCat-Flash-Thinking-2601 version is a 560 billion parameter MoE reasoning model open-sourced by Meituan and specifically designed for agents. It is the first to support an online free experience of the "Deep Thinking Mode," significantly improving complex task handling through parallel reasoning and iterative summarization. As an upgraded version of the LongCat-Flash-Thinking model, it reaches the top level among open-source models on core evaluation benchmarks such as agentic search, tool invocation, and tool-interaction reasoning. The model is now available on LongCat official website for free experience.
Main Features of LongCat-Flash-Thinking
-
Complex Reasoning: The model excels at handling complex reasoning tasks such as mathematics, logic, and programming, including automated theorem proving and solving difficult Olympiad mathematics competition problems.
-
Agentic Tool Use: The model can flexibly use external tools to assist in solving problems and improve task execution efficiency. For example, when multi-step operations are required or specific functions need to be called, it automatically selects and uses the corresponding tools.
-
General Q&A and Dialogue: It has good general language understanding and generation capabilities, can conduct fluent natural language conversations, answer questions in various fields, and provide services such as information consultation.
-
Formal Reasoning: It performs excellently in rigorous logical proofs and formal language processing, and supports applications in fields requiring precise logical verification such as mathematics and physics.
- LongCat-Flash-Thinking-2601
- Complex Task Reasoning: It can decompose and autonomously plan solutions to open-ended tasks with multiple steps and high logical complexity.
-
Powerful Tool Invocation Capabilities: It performs excellently in agentic tool invocation and interactive reasoning, possesses outstanding generalization ability, can efficiently adapt to new tools, and reduces training costs.
- Deep Analysis Mode: It supports enabling "heavy thinking," outputting more reliable solutions through parallel reasoning and iterative summarization.
Technical Principles of LongCat-Flash-Thinking
-
Mixture of Experts (MoE) Architecture: It has 560 billion parameters and dynamically activates 18.6 billion to 31.3 billion parameters according to context, achieving a balance between computational efficiency and performance.
-
Two-Stage Training Process：
-
Long Chain-of-Thought Cold-Start Training: Through a curriculum learning strategy and an SFT stage with reasoning-intensive data, it cultivates the model's basic reasoning ability.
-
Large-Scale Reinforcement Learning: Based on the DORA system for asynchronous reinforcement learning, it adopts a domain-parallel training scheme, decoupling the optimization processes of different domains before fusing them, improving the model's reasoning ability and generalization performance.
-
-
DORA System: As reinforcement learning infrastructure, it adopts a streaming architecture, multi-version design, and elastic co-location strategy, supporting efficient and stable asynchronous training on tens of thousands of accelerators, improving training efficiency and model performance.
LongCat-Flash-Thinking Project Address
- GitHub Repository：https://github.com/meituan-longcat/LongCat-Flash-Thinking
- HuggingFace Model Library：
- https://huggingface.co/meituan-longcat/LongCat-Flash-Thinking
- https://huggingface.co/meituan-longcat/LongCat-Flash-Thinking-2601
- Technical Paper：https://github.com/meituan-longcat/LongCat-Flash-Thinking/blob/main/tech_report.pdf
Application Scenarios of LongCat-Flash-Thinking
-
Software Development: In software development, it automatically generates code snippets, quickly locates and fixes errors in code, significantly improving development efficiency and code quality.
-
Scientific Research: Researchers use it to prove mathematical theorems, derive physical models, and analyze complex data, helping accelerate the scientific research process and improve research efficiency.
-
Enterprise Decision-Making: In enterprise environments, it analyzes market data, provides strategic planning and investment advice, helping enterprises make wiser decisions and enhance competitiveness.
-
Education: In education, it provides personalized learning plans based on students' learning progress, answers students' questions in real time, assists teachers in generating teaching resources, and improves teaching effectiveness.
-
Healthcare: In the healthcare field, it assists doctors in medical diagnosis, manages medical knowledge bases, provides patients with personalized health management advice, and improves the quality and efficiency of medical services.