Introduction
MiMo V2.6 Pro API is Xiaomi’s flagship for complex projects and research: combine a protocol, source documents and media within 1,048,576 input tokens, then request a text analysis of up to 131,072 tokens.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Input budget | 1,048,576 tokens |
| max_tokens ceiling | 131,072 |
| Inputs | text / images / video / audio |
Key features & highlights
The larger model for complex projects
Xiaomi positions Pro for complex multimodal projects and research. Pick it when a task has many moving parts: an experiment plan to critique, a codebase to reason about, or a report built from many sources. It costs more per token than Flash, so keep the routine volume on the smaller model.
A million tokens, four input types
Pro returns a text analysis from text, image_url, video_url, audio_url or input_audio material; images have an 8,192-token cap, input_audio uses wav or mp3, and each review request is limited to 256 messages, 128 function definitions and 4 stop strings.
Reasoning on by default
Pro defaults to high for sustained analysis; minimal, low, medium and none let you test how much reasoning your particular research question needs. Allow Pro enough max_tokens for its reasoning and final critique together: the default is 4,096 and the ceiling 131,072, so inspect finish_reason before accepting a review.
Agent loops with tools
Give Pro up to 128 OpenAI-format function definitions for evidence gathering; execution and permission checks belong to your research application. In Pro’s tool step, retain the assistant’s reasoning_content and call IDs, execute the proposed check after validation, then return its findings for the next review turn.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | For Pro research reviews, put the protocol, evidence and tool results in messages; preserve |
max_tokens | No | Allow Pro enough |
reasoning_effort | No | Pro defaults to high for sustained analysis; minimal, low, medium and none let you test how much reasoning your particular research question needs. |
stop | No | In Pro reviews, stop supports a single sequence or up to four, with matches removed; avoid a sequence that could appear inside a quotation or research finding. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Pro accepts custom temperature from 0 to 1.5 with |
top_p | No | Pro |
tools | No | Give Pro up to 128 OpenAI-format function definitions for evidence gathering; execution and permission checks belong to your research application. at most 128 items |
parallel_tool_calls | No | Use Pro |
tool_choice | No | For Pro evidence-gathering steps, validate the actual |
response_format | No | Set Pro |
Pricing
Input $0.435 / Output $0.87 per 1 million tokens
Use cases
Research and plan review
Give it a protocol, the data description and the claims, and ask for assumptions, missing controls and what would falsify each claim. Keep effort at high and set max_tokens generously.
Codebase and document analysis
For an experiment review, distinguish the hypothesis from observations, label each protocol revision and ask Pro to connect every objection to the supplied evidence. Use tools to fetch missing measurements, then return those results before asking for a revised assessment; a proposed check alone does not establish that it was performed.
Video and audio understanding
Send a product demo or an interview with its slides and ask for a structured critique or timeline. Mixed-media questions are where a larger model earns its rate.
Multi-step agents
In Pro’s tool step, retain the assistant’s reasoning_content and call IDs, execute the proposed check after validation, then return its findings for the next review turn.
Prompt guide & pro tips
Keep effort at high for hard problems
The default high is right for planning, analysis and debugging. Lower it when the answer is easy, and test on one real task before changing a whole pipeline.
Give it the whole brief
Put source material first, the question last, and ask for the output format you will validate. State which version of a document is authoritative if sources disagree.
Set a large
max_tokensand streamReasoning counts toward
max_tokens, up to 131,072. A long analysis can use the default 4,096 before the answer finishes, so raise it and stream the reply.Pass tool results and reasoning back
Append each tool result with its
tool_call_idand keep the assistant message includingreasoning_content, so the next turn continues from the same thinking.
Purpose
MiMo V2.6 Pro API is Xiaomi’s flagship for complex projects and research: combine a protocol, source documents and media within 1,048,576 input tokens, then request a text analysis of up to 131,072 tokens.
Research review
For an experiment review, distinguish the hypothesis from observations, label each protocol revision and ask Pro to connect every objection to the supplied evidence. Use tools to fetch missing measurements, then return those results before asking for a revised assessment; a proposed check alone does not establish that it was performed.
Function calls
Give Pro up to 128 OpenAI-format function definitions for evidence gathering; execution and permission checks belong to your research application.
In Pro’s tool step, retain the assistant’s reasoning_content and call IDs, execute the proposed check after validation, then return its findings for the next review turn.
Output budget
Allow Pro enough max_tokens for its reasoning and final critique together: the default is 4,096 and the ceiling 131,072, so inspect finish_reason before accepting a review.
Pro returns a text analysis from text, image_url, video_url, audio_url or input_audio material; images have an 8,192-token cap, input_audio uses wav or mp3, and each review request is limited to 256 messages, 128 function definitions and 4 stop strings.
Content policy
Research capability in MiMo V2.6 Pro does not imply unrestricted content: NSFW requests may meet model refusals, so assess suitability without assuming uncensored answers.
xiaomi/mimo-v2.6-pro/chat
Research capability in MiMo V2.6 Pro does not imply unrestricted content: NSFW requests may meet model refusals, so assess suitability without assuming uncensored answers.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "xiaomi/mimo-v2.6-pro/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.