Introduction
MiMo V2.6 Flash API brings Xiaomi’s high-frequency professional model to meeting digests, document extraction and mixed-media inboxes, with a 1,048,576-token input budget and text replies 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
Built for volume
Xiaomi positions Flash for high-volume multimodal office work, and that is how this endpoint is best used: meeting recordings, screenshots, scanned pages and long threads, one request per item. It is the lighter model of the two, so a pipeline that handles thousands of items a day is where it pays off.
Text, images, video and audio together
For a Flash digest, keep the agenda and recording references in messages; send video and audio inline within the combined 14 MB allowance, and retain tool_call_id for tool replies.
Thinking you can switch off
reasoning_effort runs from none through minimal, low and medium to high. Use none for classification, extraction and short rewrites, where it also frees temperature and top_p for you to set, and move up when a question needs working out.
Tools in the same call
Send up to 128 tool definitions, with tool_choice and parallel_tool_calls available. The model proposes a call, your code runs it, and the result goes back as a tool message, with the earlier reasoning_content kept on the assistant turn.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | For a Flash digest, keep the agenda and recording references in messages; send video and audio inline within the combined 14 MB allowance, and retain |
max_tokens | No | Flash starts with |
reasoning_effort | No | Flash offers none, minimal, low, medium and high, with high selected initially; test none on extraction before increasing effort for ambiguous notes. |
stop | No | For Flash extraction, stop accepts one sequence or up to four; choose a delimiter that cannot truncate a legitimate field, as the matched delimiter is omitted. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Flash allows temperature from 0 to 1.5 only at |
top_p | No | For Flash without reasoning, |
tools | No | Describe up to 128 functions Flash can propose for office work; your code checks arguments before updating a calendar or ticket. at most 128 items |
parallel_tool_calls | No | Enable parallel proposals in Flash when separate lookups do not depend on one another; the application still performs those lookups. Default: |
tool_choice | No | For Flash office assistants, prefer |
response_format | No | Ask Flash for text or structured JSON through |
Pricing
Input $0.14 / Output $0.28 per 1 million tokens
Use cases
Meeting and call digests
Send an audio recording or its transcript with the agenda and ask for decisions, owners and open questions as JSON. At reasoning_effort none the answer comes back quickly, which suits a queue of dozens of recordings.
Screenshot and document triage
Attach invoices, forms or UI screenshots as image parts and ask for the fields you need or the problem you can see. Short, repeated requests like these are the work Flash is positioned for.
Long-thread summaries
For repeated meeting work, label each recording and agenda item, request decisions separately from open questions, and check the resulting fields against the notes before saving them. Start with a few representative meetings at reasoning_effort=none; increase effort if the decision depends on conflicting evidence.
Tool-driven office assistants
Wire up calendar, ticket or lookup tools and let the model decide which to call. Keep effort at low or medium for routine routing and raise it for the odd ambiguous request.
Prompt guide & pro tips
Start at none or low for routine work
The default is high, which spends output tokens on reasoning. For extraction and routing, set
reasoning_effortto none or low and check quality on a handful of real items before scaling up.Mix media in one message
Make content a list of parts: text first, then
image_url,video_urloraudio_urlentries. Say in the text which part the question is about, for example the second image or the audio clip.Budget
max_tokensfor reasoningReasoning counts toward
max_tokens, so the default 4,096 can be used up before a long answer finishes at higher effort. Raise it, stream the reply and checkfinish_reasonbefore trusting a result as complete.Close the tool loop
When the model returns
tool_calls, run them, append each result as a tool message with the matchingtool_call_id, keep the assistant message as it came back, and send the conversation again.
Purpose
MiMo V2.6 Flash API brings Xiaomi’s high-frequency professional model to meeting digests, document extraction and mixed-media inboxes, with a 1,048,576-token input budget and text replies up to 131,072 tokens.
Meeting workflow
For repeated meeting work, label each recording and agenda item, request decisions separately from open questions, and check the resulting fields against the notes before saving them. Start with a few representative meetings at reasoning_effort=none; increase effort if the decision depends on conflicting evidence.
Function calls
Describe up to 128 functions Flash can propose for office work; your code checks arguments before updating a calendar or ticket.
In the Flash tool example, inspect the proposed lookup, execute it in your application and return the matching tool_call_id before requesting the digest.
Output budget
Flash starts with max_tokens at 4,096 and permits up to 131,072; reasoning uses that same allowance, so check finish_reason on long digests.
For Flash, image input has a 16,384-token allowance and the answer remains text; a request can contain 256 messages, 128 tool definitions and 4 stop sequences, with wav or mp3 accepted by input_audio.
Content policy
MiMo V2.6 Flash is offered for everyday professional tasks; its responses may refuse NSFW requests, and this page makes no uncensored-output promise.
xiaomi/mimo-v2.6-flash/chat
MiMo V2.6 Flash is offered for everyday professional tasks; its responses may refuse NSFW requests, and this page makes no uncensored-output promise.
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-flash/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.