RAG and data ยท 2026-09-22
A long-document research flow with 1M context on MiMo V2.5 Pro
How to build a research flow that analyzes a long document set in a single request with MiMo V2.5 Pro's 1 million token context window, how to add current data through your own search step, and what the model's 21 October 2026 shutdown means.
Note the shutdown date first
MiMo V2.5 Pro's catalog page states that Xiaomi shuts this model down on 21 October 2026, after which requests return a model_retired error. A successor is named: `mimo/mimo-v2.6-pro`, offered at the same price. Testing the flow in this article with `mimo/mimo-v2.6-pro` too before you put it into production avoids a surprise at the shutdown date.
Even with the shutdown date known, the model works normally until then, so the usage pattern below stays valid both right now and after migrating to the successor.
Using 1M context together with current data
The 1 million token context window lets you carry many long documents (a series of reports or contracts, for example) in a single request; but those documents only carry information up to a certain date. An analysis that needs current information has to close that gap.
The model-side `web_search` tool of MiMo models is not available on LLMTR right now; requests that add `web_search` to the `tools` array return an error. Collect current data on your side instead: fetch the relevant pages with your own search service or from sources you trust, and add their text to the context next to your documents, together with each source's address and date.
A practical flow is loading the document set and the current sources you collected into context, then asking the model to compare a claim in the documents against those sources and to state which source each conclusion rests on. Because you pick the sources, you also control which data enters the analysis.
- 1M context: carries many long documents in one request, but the information freezes at a certain date.
- The model-side web_search tool is not available on LLMTR right now; collect current data on your side.
- Add the sources you collect to the context together with their address and date.
- Ask the model to state which source each conclusion rests on.
Context size drives the cost
The main cost line in this flow is input tokens: every document and every current source you add to the context is billed again on each request. Adding only the documents and sources relevant to the question at hand, rather than filling the context for every question, keeps total cost predictable.
Frequently asked questions
Do my requests automatically go to V2.6 Pro once MiMo V2.5 Pro shuts down?
No, there is no automatic reroute; requests after shutdown return a model_retired error. You need to update the model identifier yourself before the shutdown date.
Can I use the web_search tool on MiMo V2.5 Pro?
Not on LLMTR right now; requests that include `web_search` return an error. For flows that need current information, collect the sources on your side and add them to the context as text.