Longcat-2.5-preview Operational

LongCat-2.5-Preview is a new-generation foundation model released by Meituan. It retains LongCat-2.0's 1.6T total parameters, approximately 48B active parameters, and 1M-token context, while adding native multimodal capabilities for executing long-horizon tasks across end devices, browsers, desktop software, and other scenarios. LongCat-2.5-Preview is available through the official API, supports OpenAI and Anthropic interfaces, has a maximum output of 128K tokens, and can be called directly by mainstream Agent tools such as Codex and OpenClaw.

Multimodal understanding: Adds image understanding capabilities to parse visual content, supporting cross-modal question answering, content summarization, and complex visual reasoning.
Long-horizon Agent tasks: Performs extended autonomous operations across terminals, browsers, GUIs, spreadsheets, design tools, and other scenarios.
Coding capabilities: Excels at code generation, code understanding, and automated programming tasks, continuing LongCat-2.0's Agentic Coding positioning.
Deep development-environment integration: Works efficiently with mainstream tools such as Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code.
Dual-interface access: The API supports both OpenAI and Anthropic formats, with a maximum output of 128K tokens, allowing existing Agent tools to connect directly.

MoE sparse activation architecture: Continues the mixture-of-experts approach, with 1.6 trillion total parameters but only approximately 48 billion activated for each inference. Dynamic routing calls only the expert subset most relevant to the current task, controlling computational costs while maintaining model capacity.
Native multimodality integrated into the base model: Unlike the 2.0-era approach of placing image and video capabilities in separate open-source repositories, 2.5 integrates multimodality directly into the base model, jointly modeling visual information and text within a single model. This approach inherits the DiNA (Discrete Native Autoregressive) paradigm validated by LongCat-Next, which tokenizes modalities into discrete tokens for unified modeling.
Ultra-long context support: Provides a native 1M-token context window that can accommodate very long documents, entire codebases, system logs, and complex multi-turn tasks in a single request, providing the memory foundation for multi-step execution by long-horizon Agents. This also reflects the systems-engineering achievement of supporting both massive parameter counts and million-token contexts on domestic computing clusters.
Agentic training focus: The model was trained for Agent scenarios from the outset, natively supporting tool use and multi-step reasoning, while reducing latency in long-chain tasks through inference acceleration techniques such as dlp speculative decoding.

Register an account: Open https://longcat.ai/platform/ to register and sign in to Meituan's LongCat Open Platform.
Obtain a key: Create an API Key in the console.
Install an SDK: Use the existing official OpenAI or Anthropic SDK; no additional client installation is required.
Modify the configuration: Set base_url to https://api.longcat.ai/openai (OpenAI format) or https://api.longcat.ai/anthropic (Anthropic format), then enter the key.
Specify the model: Set the model ID to LongCat-2.5-Preview and send requests using the standard OpenAI chat.completions or Anthropic messages format.
Set ANTHROPIC_BASE_URL=https://api.longcat.ai/anthropic and ANTHROPIC_AUTH_TOKEN=yourKey in the environment variables.
Set ANTHROPIC_MODEL to LongCat-2.5-Preview.
After launching Claude Code, you can call the model directly for coding tasks. OpenClaw, OpenCode, Codex, and other tools work similarly: change the base_url and model name.
Sign in on the LongCat website and select the LongCat-2.5-Preview model to chat, upload images, and try simple Agent tasks in the web interface.

Massive sparse model with low-cost inference: With 1.6T total parameters and only approximately 48B activated, it balances top-tier model capabilities with controllable inference costs.
Native million-token context: The 1M-token window can ingest an entire codebase, long documents, or complete logs at once, providing a memory foundation for long-horizon tasks.
Native multimodality in the base model: Image understanding and text are jointly modeled within the same model rather than split into separate repositories, providing a more consistent interaction experience.
Long-horizon GUI Agent capabilities: Expands from coding tasks to real software operations across terminals, browsers, spreadsheets, and design tools, covering complete workflow scenarios.
Zero-migration access through dual protocols: Supports both OpenAI and Anthropic interfaces, allowing existing Agent tools to call it directly by changing the base_url.

Entire-codebase programming assistant: Load an entire repository at once with the 1M-token context to perform cross-file refactoring, locate defects, and generate large pull requests, driven directly through Claude Code or OpenCode.
Cross-software long-horizon office automation: Have the model operate browsers, spreadsheets, and desktop software to complete unattended, multi-step tasks such as “scrape web data → clean it → enter it into Excel → generate a report.”
Enterprise log and document analysis: Feed million-token-scale system logs, contract archives, or research-paper corpora into the model as a whole for one-time deep question answering, summarization, and anomaly discovery, avoiding the context fragmentation caused by RAG chunking.
GUI testing and operations: Automatically operate terminals and desktop software to perform labor-intensive tasks such as batch configuration, UI regression testing, and automated data entry in legacy systems.
Multimodal content-production assistance: Use the new image understanding capabilities for screenshot-to-code conversion, design mockup recreation through design-tool integrations, and structured organization and repurposing of mixed image-text materials.