Meituan: LongCat 2.0

LongCat 2.0 is a sparse mixture-of-experts language model from Meituan, with 48 billion active parameters out of 1.6 trillion total. Its published role includes coding, repository-level changes, long-horizon problem solving and agent workflows. Select meituan/longcat-2.0 with POST /v1/chat/completions using model and a messages array. This profile accepts text and produces text or function-call requests within the conversation.

For a repository change, supply the relevant files or excerpts, the behaviour that should change and the constraints the implementation must preserve. Ask for a concrete result, such as a patch, a proposed change sequence or an explanation of how a failure passes through the supplied modules. For document analysis, provide the actual source text and identify the questions the answer must resolve. Label different source passages so the response can connect its conclusions to the supplied material.

The published context window is 1,048,756 tokens, and the model profile lists up to 262,144 completion tokens. Plan the source material, conversation history, tool results and requested answer within the overall context budget. Set max_tokens for the requested output allowance. A large context can hold substantial text, but the task still benefits from explicit file boundaries, the current objective and a description of what a useful answer should contain.

The supported controls include tools and tool_choice for function calling. An application can declare functions for reading a selected file, running an approved check or obtaining a record, then supply the resulting data in subsequent conversation turns. The model can request a function call; your application validates its arguments, performs the permitted action and returns the result. Keep the request, tool result and next answer together when a task spans several steps.

Other published parameters include temperature, top_p, top_k, min_p, stop, seed, frequency_penalty, presence_penalty and repetition_penalty. reasoning and include_reasoning also appear in the supported parameter list. The model profile does not support response_format, so an instruction to produce JSON remains a prompt instruction. Parse and validate any structured text before passing it into code, a database or another workflow.

Use a complete Chat Completions response when the next step needs the assembled answer, or stream text when the application displays it as it arrives. Read message content and any tool calls from the returned choices, preserving usage and response information. Assemble streamed deltas before treating a proposed patch or structured answer as complete. For a code change, evaluate the completed proposal against the supplied requirements and actual checks before applying its result.