LongCat-Flash-Chat Official Chat Template and Tool-Calling Prompt

One-sentence takeaway

Strictly assembling LongCat-Flash-Chat's round prefixes, conversation history, and <longcat_tool_call> XML wrapper according to the official template makes it possible to reproduce the multi-turn dialogue and function-calling format of its open-source weights.

Use cases

• Good for: Constructing message sequences for local/self-hosted inference, multi-turn dialogue, tool calling, and Agent harnesses.

• Not for: Treating the XML template as an OpenAI/Anthropic API request schema, or rewriting special tokens without a tokenizer configuration.

• Applicable model versions: The open-source Meituan LongCat-Flash-Chat weights; the API version has been upgraded, and the current legacy model service has been retired, so the integration surface must be confirmed first.

• Applicable clients, Agents, or APIs: The official tokenizer/chat template, SGLang, and vLLM adapters; see the repository's Deployment Guide for specific deployment parameters.

• Recommended inference tier and parameters: The model is a non-thinking foundation model; the repository does not provide a temperature/top-p combination that can be reproduced uniformly, so the deployer must fix these values.

Ready-to-use content

First-turn dialogue

[Round 0] USER:{query} ASSISTANT:

First-turn dialogue with a system prompt

SYSTEM:{system_prompt} [Round 0] USER:{query} ASSISTANT:

Multi-turn dialogue

SYSTEM:{system_prompt} [Round 0] USER:{query} ASSISTANT:{response}</longcat_s>... [Round N-1] USER:{query} ASSISTANT:{response}</longcat_s> [Round N] USER:{query} ASSISTANT:

Tool calling

{tool_description}

## Messages
SYSTEM:{system_prompt} [Round 0] USER:{query} ASSISTANT:

## Tools
You have access to the following tools:

### Tool namespace: function
#### Tool name: {func.name}
Description: {func.description}
InputSchema:
{json.dumps(func.parameters, indent=2)}

For each function call, return a JSON object inside its own XML tag:
<longcat_tool_call>
{"name": <function-name>, "arguments": <args-dict>}
</longcat_tool_call>

Testing/workflow steps

1. Read the actual chat template from the repository's tokenizer_config.json; do not rely only on the simplified example in the README.

2. First run a tool-free first turn and a multi-turn echo test to confirm that the special end marker and round concatenation match.

3. Serialize the tool schema into ## Tools, requiring the model to output only JSON wrapped in <longcat_tool_call>.

4. Have the harness parse each XML block, execute the function, return the result as the next-round USER/tool context, and then independently verify the final answer.

5. Record the model version, tokenizer, inference engine, context length, and tool-calling errors; do not mix the server API's JSON tool schema with the local template.

Original evidence and data

• The official repository provides four templates: first turn, system prompt, multi-turn, and ToolCall.

• The tool-calling convention uses <longcat_tool_call> XML tags, with the function name and parameter JSON inside; multiple function calls use multiple consecutive tags.

• The official repository says that SGLang and vLLM adapters are available, with detailed deployment instructions in docs/deployment_guide.md.

Scope boundaries

• The template defines the input/output convention for the open-source weights; compatibility with every third-party API gateway is not guaranteed.

• The legacy API service has been retired according to the official Change Log; when calling through a new platform, use the current model list and API documentation as the authority.

• The ToolCall template defines only the format. It does not guarantee correct tool selection, safe arguments, or successful completion of the final task; the execution layer must perform schema validation, permission isolation, and timeout handling.

Source excerpts or observations (for compliant short quotations only)

• The official template writes rounds as [Round N] USER... ASSISTANT and retains the <longcat_s> end marker in multi-turn exchanges.

• The official tool format requires function calls to appear inside <longcat_tool_call> tags rather than as free-text descriptions.