MiMo-V2.6-Flash Deep Thinking Configuration and Multi-turn Tool-calling Workflow

One-sentence takeaway

mimo-v2.6-flash supports toggling deep thinking with thinking.type, which is enabled by default. When it is enabled, do not customize temperature or top_p, and pass through the historical assistant messages' reasoning_content in full during multi-turn tool calls.

Use cases

• Suitable tasks: Complex reasoning, code generation, mathematical computation, and multi-step analysis.

• Unsuitable tasks: Simple Q&A where response latency or output budget is especially sensitive and deep analysis is unnecessary; the official documentation states that enabling deep thinking increases latency.

• Applicable model version: mimo-v2.6-flash. The source also lists mimo-v2.6-pro, mimo-v2.6-pro-ultraspeed, mimo-v2.5-pro, and mimo-v2.5; this article does not treat example responses for those versions as Flash data.

• Applicable client, Agent, or API: The OpenAI Chat Completions-compatible interface of the MiMo API; in the Python SDK, thinking is not a standard OpenAI parameter and must be placed in extra_body.

• Recommended reasoning tier and parameters: The documentation does not provide a Flash-specific reasoning tier. thinking.type uses enabled or disabled; in deep thinking mode, the effective values are temperature=1.0 and top_p=0.95, and custom values are not supported.

Ready-to-use content

The following is a configuration skeleton with the example model ID replaced by mimo-v2.6-flash according to the official supported-model list. The Python, response, and streaming examples in the source all use mimo-v2.6-pro, so this code expresses Flash's officially supported scope and general parameter rules, rather than an officially provided Flash test response.

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-flash",
    messages=[
        {"role": "user", "content": "Break down this multi-step problem and provide verifiable conclusions."}
    ],
    max_completion_tokens=2048,
    stream=True,
    extra_body={"thinking": {"type": "enabled"}},
)

for chunk in completion:
    print(chunk.model_dump_json())

To turn off deep thinking, simply change it to:

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

Testing or workflow steps

1. Create a Chat Completions request using https://api.xiaomimimo.com/v1 and mimo-v2.6-flash.

2. In the Python SDK, put thinking in extra_body, and set thinking.type to enabled or disabled; Flash defaults to enabled.

3. In deep thinking mode, do not rely on custom temperature or top_p; even if they are passed, the effective values are forced to 1.0 and 0.95.

4. Set max_completion_tokens according to the total output budget. It consumes tokens for both the thinking content and the final answer; when the thinking is lengthy, the answer may be truncated.

5. Set stream: true when you need to observe the process in real time. The official documentation states that streaming responses first return the thinking content progressively through reasoning_content, then return the final answer through content.

6. In multi-turn Agent tool calls, put the reasoning_content from historical assistant messages back into the subsequent request in full; the API returns 400 if it is missing.

7. Regression checks should cover at least: both enabled and disabled configurations, the order of streaming fields, a sufficient output budget, and the next request with tool-call history.

Original evidence and data

• The official “Supported Models” section explicitly lists mimo-v2.6-flash, along with Pro, Pro UltraSpeed, V2.5 Pro, and V2.5; the list does not provide Flash-specific performance, latency, or token test data.

• The official “Request Parameters” section specifies thinking.type, with enabled or disabled as its values; the models listed as enabled by default on the page include mimo-v2.6-flash.

• The official “Parameter Limitations” section states that the supported models above do not support custom temperature or top_p in deep thinking mode, and that the effective values are forced to 1.0 and 0.95.

• The official “Multi-turn Conversation Pass-through Requirements” section states that when Agent multi-turn history contains tool calls, subsequent requests containing tool calls must pass through reasoning_content in full; otherwise, the API returns 400. Missing this field may also result in incomplete context, weaker instruction following, and more hallucinations.

• The “Other Notes” section on the page states that max_completion_tokens limits both the thinking content and the final answer, and recommends providing a sufficient budget when the thinking is lengthy; enabling deep thinking increases latency, so stream: true is recommended.

• The Agent products listed as affected on the page include: under the OpenAI-compatible protocol, TRAE, Cursor, Roo Code, Codex, GitHub Copilot CLI, Zed, AutoGen, and Goose; under the Anthropic-compatible protocol, TRAE, GitHub Copilot CLI, AutoGen, Goose, OpenClaw, OpenCode, and Kilo Code.

• The complete Python, Curl, response, and multi-turn tool-call examples on the page all set model to mimo-v2.6-pro; these Pro examples are not treated as Flash test results.

Scope and limitations

• “Flash is included in the supported list” only proves that the documentation declares support; this article has not independently measured Flash API requests, accuracy, latency, or token usage.

• reasoning_content is an interface field; whether raw thinking content should be shown to end users should be decided separately by the product based on privacy, security, and interaction requirements.

• If max_completion_tokens is too small, lengthy thinking will consume the space available for the final answer; set the budget according to task complexity and answer length rather than copying the example's 1024.

• Pro, Pro UltraSpeed, V2.5 Pro, and V2.5 are used only to describe the documentation's shared supported scope; this does not support inferring that Flash has the same quality, speed, or cost as those models.

Reproduction notes

During collection, Tabbit was used to fully open and read the official page, whose page update date was 2026-09-22. To reproduce the process, prepare a MiMo API key and send separate enabled, disabled, and stream=True requests using the Flash configuration above; for multi-turn tool calls, also append each round's assistant message, including the complete reasoning_content, back to messages. This article does not claim that these requests were executed locally, nor does it treat the official Pro examples as Flash test data.