Longcat-2.5-preview - Meituan's New Generation Large Model

What is Longcat-2.5-preview

LongCat-2.5-Preview is Meituan's new generation large model, continuing LongCat-2.0 's 1.6T total parameters, approximately 48B activated parameters and 1M token context, with newly added native multimodal capabilities, capable of executing long-process tasks for terminals, browsers, desktop software and other scenarios. LongCat-2.5-Preview is now connected to the official API, compatible with OpenAI and Anthropic interfaces, with a maximum output of 128K tokens, supporting direct invocation by mainstream Agent tools such as Codex、OpenClaw .

Main Features of Longcat-2.5-preview

• Multimodal Understanding: Newly added image understanding capability, able to parse image content, supporting cross-modal question answering, content summarization and complex visual reasoning.
• Long-range Agent Tasks: Executes long-process autonomous operations for terminals, browsers, GUI, spreadsheets, design tools and other scenarios.
• Coding Capability: Outstanding performance in code generation, code understanding and automated programming tasks, continuing the Agentic Coding positioning of LongCat-2.0.
• Deep Adaptation to Development Environments: Efficient collaboration with mainstream tools such as Claude Code、Hermes、OpenClaw、OpenCode、Kilo Code .
• Dual-interface Openness: The API is simultaneously compatible with OpenAI and Anthropic formats, with a maximum output of 128K tokens, and existing Agent tools can be directly connected.

Technical Principles of Longcat-2.5-preview

• MoE Sparse Activation Architecture: Continuing the mixture-of-experts approach, with 1.6 trillion total parameters but only approximately 48 billion parameters activated per inference, dynamically routing to call only the subset of experts most relevant to the current task, controlling computational cost while ensuring model capacity.
• Native Multimodality Integrated into the Base: Unlike the 2.0 era, which split images and videos into independent open-source repositories, 2.5 directly integrates multimodal capabilities into the base itself, enabling visual information and text to be jointly modeled within the same model—this approach inherits from the DiNA (Discrete Native Autoregressive) paradigm validated by LongCat-Next, which lexicalizes modalities into discrete tokens for unified modeling.
• Ultra-long Context Support: Native 1M token context window, capable of accommodating ultra-long documents, entire codebases, system logs and multi-turn complex tasks in a single pass, providing the memory foundation for multi-step execution of long-range Agents; this is also a systems engineering achievement of simultaneously supporting ultra-large parameters and million-level context on domestic computing power clusters.
• Agentic Training Positioning: The model was trained for Agent scenarios from the outset, natively supporting tool invocation and multi-step reasoning, and reducing latency in long-chain tasks through inference acceleration technologies such as dlp speculative decoding.

How to Use Longcat-2.5-preview

• Register an Account: Open https://longcat.ai/platform/ to register and log in to the Meituan LongCat open platform.
• Obtain a Key: Create an API Key in the console.
• Install the SDK: Use the existing official OpenAI or Anthropic SDK, no additional new client installation is required.
• Modify the Configuration: Change base_url to https://api.longcat.ai/openai(OpenAI format) or https://api.longcat.ai/anthropic(Anthropic format), and fill in the Key.
• Specify the Model: Fill in the model ID as LongCat-2.5-Preview, and send requests in the standard OpenAI chat.completions or Anthropic messages format.

• Set in the environment variables ANTHROPIC_BASE_URL=https://api.longcat.ai/anthropic and ANTHROPIC_AUTH_TOKEN=你的Key。
• Set ANTHROPIC_MODEL to LongCat-2.5-Preview。
• After starting Claude Code, you can directly call this model to perform coding tasks; the same applies to tools such as OpenClaw, OpenCode, and Codex—just change the base_url and model name.

• Visit LongCat Official website, log in, and select the LongCat-2.5-Preview model to experience chat, image upload, and simple Agent tasks on the web.

Longcat-2.5-preview's core advantages

• Ultra-large sparse model + low-cost inference: 1.6T total parameters with only about 48B activated, balancing top-tier model capabilities with controllable inference costs.
• Million-level native context: A 1M token window can swallow an entire codebase, long documents, or complete logs in one go, providing a memory foundation for long-horizon tasks.
• Native multimodality integrated into the base: Image understanding and text are jointly modeled within the same model, and are no longer split into a separate repository, making the interaction experience more unified.
• Long-process GUI Agent capabilities: Expanding from coding tasks to real software operations in terminals, browsers, spreadsheets, and design tools, covering complete workflow scenarios.
• Dual-protocol zero-migration access: Compatible with both OpenAI and Anthropic interfaces; existing Agent tools can call it directly by changing the base_url.

Comparison of Longcat-2.5-preview with similar competitors

Application scenarios of Longcat-2.5-preview

• Full-codebase-level programming assistant: Use the 1M context to load an entire repository at once, completing cross-file refactoring, defect localization, and large PR generation, directly driven through Claude Code / OpenCode.
• Cross-software long-process office automation: Let the model operate browsers + spreadsheets + desktop software to complete unattended multi-step tasks such as "scrape web data → clean → fill into Excel → generate report".
• Enterprise-level log and document analysis: Feed in million-token-scale system logs, contract archives, or paper corpora as a whole to perform one-shot deep question answering, summarization, and anomaly mining, avoiding the context fragmentation caused by RAG chunking.
• GUI interface testing and operations: Automatically operate terminals and desktop software to perform labor-intensive operations such as batch configuration, interface regression testing, and automated data entry for legacy systems.
• Multimodal content production assistance: Combined with the newly added image understanding capability, perform screenshot-to-code conversion, design mockup reproduction (integrated with design tools), and structured organization and re-creation of mixed image-text materials.