Xiaomi MiMo-V2.6-Pro Official API Integration and Reasoning Configuration

One-sentence conclusion

This official configuration can be used to connect to mimo-v2.6-pro through the OpenAI-compatible protocol, with deep thinking, streaming output, and multi-turn tool calls enabled as needed.

Applicable scenarios

• Suitable tasks: Complex projects, long-horizon tasks, high-value work, cybersecurity and research needs, and Agents that require multi-step reasoning, tool calls, structured output, or full-modality understanding.

• Unsuitable tasks: Highly real-time interactions that are extremely latency-sensitive; for these scenarios, the official documentation recommends evaluating mimo-v2.6-pro-ultraspeed. For frequent, large-scale office tasks where cost matters more, mimo-v2.6-flash may be evaluated.

• Applicable model version: mimo-v2.6-pro; this article does not generalize the configuration to other MiMo versions.

• Applicable clients, Agents, or APIs: OpenAI Chat Completions API and Anthropic Messages API. The official documentation also lists TRAE, Cursor, Roo Code, Codex, GitHub Copilot CLI, Zed, AutoGen, and Goose under the OpenAI-compatible protocol, and TRAE, GitHub Copilot CLI, AutoGen, Goose, OpenClaw, OpenCode, and Kilo Code under the Anthropic-compatible protocol.

• Recommended reasoning mode and parameters: Use thinking.type: enabled for complex reasoning or Agent tasks; disabled can be used for ordinary short answers. For non-thinking calls, the official example uses temperature=1.0 and top_p=0.95. For deep thinking, do not rely on custom temperature or top_p; the official documentation says the recommended defaults of 1.0 and 0.95 are used in practice. max_completion_tokens should cover the combined length of the reasoning content and final answer.

Ready-to-use content

The following is a minimal runnable OpenAI Python SDK configuration organized from the official pages. Read the API key only from an environment variable; do not put a real key in code or commit it to a repository.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MIMO_API_KEY"],
    base_url="https://api.xiaomimimo.com/v1",
)

completion = client.chat.completions.create(
    model="mimo-v2.6-pro",
    messages=[
        {
            "role": "system",
            "content": (
                "You are MiMo, an AI assistant developed by Xiaomi. "
                "Today's date: {date} {week}. "
                "Your knowledge cutoff date is December 2024."
            ),
        },
        {"role": "user", "content": "Complete the task. First list the key assumptions, then provide verifiable conclusions."},
    ],
    max_completion_tokens=4096,
    temperature=1.0,
    top_p=0.95,
    stream=False,
    extra_body={"thinking": {"type": "enabled"}},
)

print(completion.model_dump_json())

The official documentation provides the following Chinese system prompt:

You are MiMo (the Chinese name is also MiMo), an AI assistant developed by Xiaomi.
Today's date: {date} {week}. Your knowledge cutoff date is December 2024 (month 12).

The official page also provides an English version:

You are MiMo, an AI assistant developed by Xiaomi.
Today's date: {date} {week}. Your knowledge cutoff date is December 2024.

If you use the Anthropic Messages API, adapt the configuration to the official endpoint as follows:

import os
from anthropic import Anthropic

client = Anthropic(
    api_key=os.environ["MIMO_API_KEY"],
    base_url="https://api.xiaomimimo.com/anthropic",
)

message = client.messages.create(
    model="mimo-v2.6-pro",
    max_tokens=4096,
    system="You are MiMo, an AI assistant developed by Xiaomi. "
           "Today's date: {date} {week}. Your knowledge cutoff date is December 2024.",
    messages=[{"role": "user", "content": "Complete the task and output verifiable steps."}],
    top_p=0.95,
    temperature=1.0,
    stream=False,
)

These code samples combine the official API integration and parameter configuration; they are not Xiaomi-published task-specific prompts. The user task itself should still add inputs, acceptance criteria, and an output format appropriate to its goal.

Test/workflow steps

1. Create an API key on the Xiaomi MiMo API platform and set the environment variable:

   export MIMO_API_KEY='sk-xxxxx'
   pip install -U openai

2. First send a short request with thinking.type: disabled to confirm that the Base URL, key, and model ID are correct.

3. For complex reasoning, code, mathematics, or multi-step Agent tasks, change it to thinking.type: enabled, and set max_completion_tokens to cover the total budget for reasoning and the answer.

4. To enable streaming output, set stream to True; the official documentation says that reasoning content is returned incrementally through reasoning_content first, followed by the final content.

5. If tool calls are enabled and the conversation continues over multiple turns, save and return the assistant's reasoning_content unchanged on every turn. When historical turns contain tool calls but this field is omitted, the official documentation says the API may return 400 and instruction following may worsen or hallucinations may increase.

6. Record the model, reasoning toggle, max_completion_tokens, input and output tokens, and whether tool calls succeeded for each request; then compare latency and cost against the same task with reasoning disabled.

Original evidence and data

Model capabilities and quotas

The official “Model list” page lists the following for the text-generation model:

• Model ID: mimo-v2.6-pro

• Capabilities: Text generation, full-modality understanding, deep thinking, streaming output, function calling, structured output, and web search

• Context window: 1M

• Maximum output: 128K

• Rate limits: maximum RPM 100, maximum TPM 10M

• Quick selection: mimo-v2.6-pro is recommended for complex projects, long-horizon tasks, high-value work, cybersecurity, and research needs

The same page positions mimo-v2.6-pro-ultraspeed for highly real-time scenarios and mimo-v2.6-flash for frequent, large-scale professional office tasks. These are the official selection descriptions and do not represent independent performance conclusions about adjacent models.

Endpoints and model ID

The “First API call” page states that the platform is compatible with the OpenAI API and Anthropic API, and provides:

• OpenAI real-time inference Base URL: https://api.xiaomimimo.com/v1

• Anthropic real-time inference Base URL: https://api.xiaomimimo.com/anthropic

• Model used in the OpenAI SDK call: mimo-v2.6-pro

• Model used in the Anthropic SDK call: mimo-v2.6-pro

• Default parameters in the official example: max_completion_tokens=1024, temperature=1.0, top_p=0.95, stream=False

This article raises max_completion_tokens to 4096 in its examples to leave a more practical total budget for deep thinking and the final answer; this is a configuration recommendation, not an official fixed value.

Deep thinking

The “Deep thinking” page explicitly states that mimo-v2.6-pro supports thinking.type and enables it by default; enabled or disabled can be used to control it. The official documentation also states:

• Deep thinking is suitable for complex reasoning, code generation, mathematical calculations, and multi-step analysis.

• Custom temperature and top_p are not supported during deep thinking; even if supplied, the actual values remain 1.0 and 0.95.

• max_completion_tokens limits both the reasoning content and the final answer.

• Enabling thinking increases response latency; the official documentation recommends combining complex tasks with stream: true to view output in real time.

• During multi-turn tool calls, the reasoning_content in the assistant message containing the tool call must be returned in full.

The key fragment of the official Python parameter example is:

extra_body={
    "thinking": {"type": "enabled"}
}

To disable thinking, the official example changes it to:

extra_body={
    "thinking": {"type": "disabled"}
}

Scope and limitations

• This article only establishes that the official documentation publishes the integration method, parameters, and model capabilities; it does not establish independent results for mimo-v2.6-pro on any particular business task.

• The 1M context window and 128K maximum output listed on the official model page are service limits; they do not mean that every task will consistently receive the same quality or latency.

• During deep thinking, do not treat temperature and top_p as effective tuning controls; prioritize adjusting the prompt, tool definitions, context, and token budget.

• Multi-turn Agents need to persist reasoning_content; saving only the final text and tool results may cause subsequent requests to fail or quality to decline.

• The date in the official system prompt is a placeholder; replace it with the actual date and weekday for real calls rather than copying the example date after it becomes outdated.

• The official documentation lists a knowledge cutoff of December 2024 (month 12); verify time-sensitive information through web search tools, external retrieval, or materials provided by the user.

• The code samples contain no real key, tool server, or business data. Before deployment, add timeouts, retries, backoff, log redaction, and tool permission controls.

Source excerpts or observations

1. Model list (official, updated 2026-09-21): The page directly lists mimo-v2.6-pro, a 1M context window, a 128K maximum output, text/full-modality/deep-thinking/function-calling/structured-output/web-search capabilities, and complex projects, long-horizon tasks, high-value work, cybersecurity, and research needs as recommended scenarios.

2. First API call (official, updated 2026-09-22): The page directly provides OpenAI and Anthropic Base URLs, the mimo-v2.6-pro model ID, Python SDK examples, and the official Chinese and English system prompts.

3. Deep thinking (official): The page directly states that Pro supports thinking.type and enables deep thinking by default; it also gives the extra_body syntax, parameter limitations, streaming behavior, and the requirement to return reasoning_content during multi-turn tool calls.

4. Model release log (official): The page describes mimo-v2.6-pro as a full-modality, ultra-high-performance, trillion-parameter flagship reasoning model for complex projects, long-horizon tasks, high-value work, cybersecurity, and research needs. This is the vendor's positioning and should not be treated on its own as a third-party evaluation score.