MiMo API: Xiaomi's Long-Context AI Models

Direct pay-as-you-go price

Xiaomi publishes both China-region CNY and overseas USD prices. The overseas figures below were checked on September 11, 2026 and are per 1M tokens:

The China-region prices are ¥0.025, ¥3.00, and ¥6.00 respectively. These are parallel regional price lists, not values derived by exchange-rate conversion.

Xiaomi currently describes cache writes as temporarily uncharged. “Temporarily” is an operational warning: keep that meter in your cost model and recheck the rate page before a large rollout. Web search is billed separately from the token prices above.

Direct curl request

curl https://api.xiaomimimo.com/v1/chat/completions \
  -H "Authorization: Bearer $MIMO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mimo-v2.5-pro","messages":[{"role":"user","content":"Extract the three decisions in this document."}]}'

Python with an OpenAI client

import os
from openai import OpenAI

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

result = client.chat.completions.create(
    model="mimo-v2.5-pro",
    messages=[{"role": "user", "content": "Return a JSON migration plan."}],
)
print(result.choices[0].message.content)

When routing through AIWave, use AIWave's catalog ID: xiaomi/mimo-v2.5-pro. The provider prefix is part of that gateway ID; do not silently reuse the direct ID. The base URL becomes https://aiwave.live/v1, and the same OpenAI client structure can be used with an AIWave key.

Check AIWave's dated USD price before calling. Its public rate is the price of the AIWave route and billing layer, not Xiaomi's direct rate. The operational value is a common key and invoice across providers; the cost is the gateway markup and another component to monitor.

### Structured output and tool safety

MiMo lists structured output and tool calls, but production readiness depends on your schema. Test required fields, enums, nested objects, and malformed tool arguments. Validate every tool call before execution and separate read-only tools from actions that change external state. If a timeout occurs after an action may have completed, reconcile the external system before retrying. A model's tool-call capability does not make the underlying operation idempotent.

How to spend a million-token window well

Start with a token budget, not a file dump. For a codebase, send the repository map, relevant interfaces, the current change, and failing evidence. Retrieve additional files when the model identifies a real dependency. For a document workflow, separate source text from instructions and reserve enough output for citations or structured results.

Keep stable material in a deterministic prefix if you expect cache hits. Moving timestamps or request IDs to the front can invalidate a large cached region. Log the provider-reported cached input instead of estimating it from repeated characters.

The public limits also deserve load testing. Xiaomi lists RPM 100 and TPM 10M for the model, but your account, region, or contract can differ. Run a small concurrency test, cap retries, and avoid retrying a tool action that has already changed external state.

For production, store the returned model ID, token usage, latency, HTTP status, and charge with the request ID. A 1M context window can remove a retrieval step; it can also create a much larger failure if timeouts and output limits are left implicit.

Primary references

• MiMo V2.5 Pro model page
• MiMo pay-as-you-go pricing
• MiMo model updates
• MiMo V2.5 open-source announcement