MiniMax M2.7 Official Default Prompt and XML Tool-Calling Template

One-sentence takeaway

When self-hosting MiniMax M2.7, start with the official recommended temperature=1.0, top_p=0.95, top_k=40 and default system prompt, and define/parse tools using its <minimax:tool_call> XML syntax.

Use cases

• Suitable tasks: Self-hosted Agents, OpenAI SDK-compatible services, coding/research tool calls, and multi-turn engineering workflows.

• Unsuitable tasks: Passing raw XML output directly to production tools; strict parsing, schema validation, and permission checks are required first.

• Applicable model version: MiniMax-M2.7; the repository says M2.7 uses the same tool-call syntax as M2.

• Applicable clients, Agents, or APIs: vLLM, SGLang, Transformers, and OpenAI Python SDK-compatible endpoints; cloud providers may customize the parser/configuration.

• Recommended inference tier and parameters: The official repository recommends temperature=1.0, top_p=0.95, and top_k=40; do not arbitrarily lower temperature to 0 without testing.

Ready-to-use content

Official default system prompt:

You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax.

Official XML structure for tool calls (replace the tool name, parameter names, and parameter values according to the actual schema):

<minimax:tool_call>
<invoke name="search_web">
<parameter name="query_tag">["technology", "events"]</parameter>
<parameter name="query_list">["OpenAI latest release"]</parameter>
</invoke>
</minimax:tool_call>

Skeleton for an OpenAI SDK-compatible tool definition:

tools = [{
    "type": "function",
    "function": {
        "name": "search_web",
        "description": "Search approved sources and return source URLs.",
        "parameters": {
            "type": "object",
            "properties": {
                "query": {"type": "string", "description": "One precise search query."}
            },
            "required": ["query"]
        }
    }
}]

Test/workflow steps

1. Start the same inference server with the official recommended parameters, and confirm that the chat template matches the tokenizer version.

2. Have the model make a weather/search tool call, save the complete raw output, and check whether there is one or more <invoke> blocks.

3. Prefer the built-in parser in vLLM/SGLang; if parsing manually, parse the XML block, invoke, and parameter layers in sequence, then convert types according to the JSON Schema.

4. Before executing a tool, validate its name, parameter types, required fields, permissions, and allowlist; afterward, return the structured result as a tool message.

5. For the same tool set, compare the call success rate across temperature/top-p configurations, different providers, and streaming/non-streaming modes.

Raw evidence and data

• The official repository explicitly recommends temperature=1.0, top_p=0.95, and top_k=40, and provides the default system prompt.

• The official tool guide recommends using vLLM or SGLang for tool-call parsing; other frameworks require manually parsing the XML.

• The official output format is <minimax:tool_call> → <invoke name> → <parameter name>, and multiple invokes are supported in one block.

• Tool parameters are defined with JSON Schema; the official example requires name, description, parameters.properties, required, and type.

Applicability boundaries

• These parameters are official starting values for self-hosting, not necessarily optimal for every task or cloud provider; use task-level evaluations.

• Raw XML may contain formatting errors, missing quotation marks, or escaping issues; do not extract it with a regex and execute it directly.

• Chat templates, tool parsers, and sampling configurations may differ across inference engines/providers; the community has reported stripped double quotation marks, so provider A/B testing is required.

• The default system prompt does not include business permissions, data privacy, or completion acceptance criteria; production systems must add these separately.

Source excerpt or observation (compliance-short quotation only)

The official tool guide says “strongly recommend using vLLM or SGLang for parsing tool calls” (compliance-short quotation).