What is Xiaomi MiMo V2.5 Pro?
Xiaomi MiMo V2.5 Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion active parameters, released as open weights under a permissive licence and free to download and run. The architecture interleaves local sliding-window attention with full attention at a ratio of six to one and a window of 128 tokens, a design the vendor says cuts KV-cache storage by nearly seven times at long context while holding quality through a learnable attention-sink bias. A multi-token-prediction module built as a small dense feed-forward block is credited with roughly tripling output throughput. Pre-training ran over 27 trillion tokens in FP8 mixed precision at a native 32K sequence length before the window was extended, and post-training moved through supervised fine-tuning, domain specialisation and multi-teacher on-policy distillation. Two relationships are worth fixing before anything else. First, the retirement is unconditional: the notice names 21 October 2026 at 10:00 Beijing time, states that no system replacement model exists, and says requests under the old name will error afterwards. A second table on the same page shows that an earlier retirement round did name a replacement, so the absence here is a decision rather than a formatting gap. Second, the window that is documented and the window that is displayed here are not the same thing. The specification block reads 1M tokens, and this interface shows no context figure at all, because the catalogue it reads carries no context field for this model.
What Xiaomi MiMo V2.5 Pro is worth using for
A 1M-token window, and a chat box that shows no number for it
The documented context window is one million tokens and the maximum output is 128K tokens. Neither figure appears in the workspace above. This site reads a catalogue that carries no context or maximum-output field for this model, so the composer falls back to its own limit, each reply is requested with at most 4,096 output tokens, and no context figure is rendered anywhere. That gap matters more on this page than on one where the two sides agree, because the vendor’s own long-context evidence is unusually specific. On the GraphWalks benchmark the model scores 0.56 on the breadth-first track and 0.92 on the parent-lookup track at a 512K input, and 0.37 and 0.62 at the full million, while the previous generation MiMo-V2-Pro collapses to 0.00 at the same length. A page that repeated only the window size would hide both the graceful degradation and the local ceiling.
I am going to paste a set of long files and then ask questions across all of them. Read everything before answering anything. For each answer, name the file you relied on, say which part of my question the material did not settle, and tell me where you would have needed a file I did not provide.
Thinking you can switch off, and a switch that is greyed out here
This is where the model differs from most of the reasoning models this site covers. Deep thinking is controlled by a request parameter, thinking.type, which accepts enabled or disabled, and it is on by default for this model. Turning it off is a supported configuration rather than an unsupported one, and it changes the sampling rules in the same move: while thinking is enabled the model does not accept a custom temperature or top_p, and both are forced to 1.0 and 0.95, so a prompt that tunes them is tuning values the model ignores. The workspace above does not expose the switch at all. The catalogue this site reads publishes no reasoning field for this model, so the thinking control is present and unavailable, which means the one setting most worth testing here cannot be tested from this page.
Answer first without showing me any intermediate reasoning, and give me a direct recommendation plus the two objections a careful reviewer would raise. Then, when I ask, redo the same question with your reasoning shown step by step so I can compare the two answers.
A multi-turn contract that returns a 400 when it is broken
Deep thinking imposes a rule that decides how a long session has to be run. When an agent carries tool calls across several turns, the full reasoning_content from every earlier assistant message has to be sent back with the request; the guide states that omitting it returns a 400 error, and adds that a missing reasoning field can leave the context incomplete, with weaker instruction following and more fabrication as the result. Two smaller details come from the same page. The thinking block is not a standard OpenAI parameter, so a request built with the OpenAI SDK has to place it inside extra_body. And because thinking content is billed and streamed before the answer, a streaming client receives reasoning_content first and the answer afterwards, which is what makes the added latency visible rather than silent. None of this is reachable from the page you are reading, because the switch itself is unavailable above.
Keep the reasoning from this session available so that follow-up questions can build on it. When I send a tool result back, treat the earlier thinking as part of the context rather than as something to be summarized away, and tell me whenever something I omit would break that chain.
Open weights, and the vendor’s own case studies
The release is open and the licence is permissive: weights, tokenizer and model card are on Hugging Face under an MIT licence, and the model card carries deployment guidance for SGLang and vLLM, so the model can be run outside any hosted service. The vendor also publishes three case studies with numbers attached. A Rust compiler written from scratch for a university course project finished in 4.3 hours across 672 tool calls, scoring 233 out of 233, with intermediate readings of 137 of 233 on the first compile, 110 of 110 on the intermediate representation and 103 of 103 on the back end. A full video editor took 11.5 hours and 1,868 tool calls and produced 8,192 lines of code. An analogue circuit design closed in about an hour, improving line regulation from 0.65 to 0.03 millivolts per volt. On token efficiency the vendor reports 64 per cent pass-cubed on ClawEval at roughly 70,000 tokens per trajectory, which it says is 40 to 60 per cent fewer tokens than three named competitors at comparable capability. These are the vendor’s own measurements, and this page treats them as exactly that.
Take this task end to end without asking me to confirm each step, but write the plan first and keep a running log of the tool calls you make. When you finish, show me the log and tell me which steps you would skip if you ran it again.
Official Xiaomi MiMo V2.5 Pro specifications
- Identifier
- mimo-v2.5-pro
- Parameters
- 1.02 trillion total, 42 billion active — mixture of experts
- Architecture
- Sliding-window and global attention interleaved 6:1 with a 128-token window, plus multi-token prediction
- Context window
- 1M tokens
- Maximum output
- 128K tokens
- Thinking
- thinking.type accepts enabled or disabled; on by default for this model
- Sampling parameters
- temperature 1.0 and top_p 0.95 — both forced while thinking is enabled, and neither can be customised then
- Pre-training
- 27 trillion tokens in FP8 mixed precision, native 32K sequence length extended to 1M
- Licence
- MIT — weights, tokenizer and model card on Hugging Face
- API protocols
- OpenAI and Anthropic compatible from https://api.xiaomimimo.com/v1
Xiaomi MiMo V2.5 Pro and Claude Opus 4.6 on the vendor’s own table
Every value below is read from a single page — the full benchmark table on the Xiaomi MiMo V2.5 Pro product page, which prints this model and seven comparison models in adjacent columns under one evaluation setup. Two columns are reproduced here, this model’s and Claude Opus 4.6’s, and every row carries a value on both sides, so one official artefact accounts for the whole table. The rows that go against the model are kept rather than filtered out, and they are the majority: of the eight measures below, Xiaomi MiMo V2.5 Pro leads on two and trails on six. It wins the multi-turn agent benchmark and the terminal benchmark; it trails on a general-agent Elo rating, a task-completion evaluation, a graduate-level science exam, two software-engineering sets and an implementation-ranked coding measure. These are the vendor’s own readings of its own model against a competitor, not an independent measurement, and no score here predicts the quality of your conversation.
| Official benchmark | Xiaomi MiMo V2.5 Pro | Claude Opus 4.6 |
|---|---|---|
| GDPVal-AA · Elo · Agentic | 1,581 | 1,606 |
| τ³-bench · Agentic | 72.9 | 72.4 |
| Claw-Eval · pass^3 · Agentic | 63.8 | 70.4 |
| Humanity's Last Exam · Reasoning | 48.0 | 53.0 |
| SWE-Bench Pro · Coding | 57.2 | 57.3 |
| SWE-bench Verified · Coding | 78.9 | 80.8 |
| Terminal-Bench 2.0 · Coding | 68.4 | 65.4 |
| FrontierSWE · Implementation rank · Coding | #3.4 | #2.0 |
How to use Xiaomi MiMo V2.5 Pro online
Start from the end date, not the specification
The single most consequential fact about this model is that it has a published expiry. The official notice puts it at 21 October 2026 at 10:00 Beijing time, states that no system replacement model exists, and says that requests using the old name will receive an error after that moment. A second table on the same page shows what a handled retirement looks like: the earlier round moved users from one generation to the next automatically. Nothing here says what to migrate to, because the vendor does not say either. Read that as a planning constraint rather than a warning, and check the notice again before you build anything that depends on this model name.
Decide which of the three sources you are trusting
Three different artefacts describe this model and they answer different questions. The specification block says what the API supports, and its window and ceiling are the ones quoted in the cards above. The pricing page says what a token costs and which features are excluded. The product page is where the architecture claims, the case studies and the benchmark table live, and its numbers are the vendor’s own evaluation of its own model. Where two of those disagree, this page says so instead of picking one. The one place to be careful is the feature list: the specification block records this model’s input modality as text, so this page does not repeat the broader multimodal phrasing that appears elsewhere on the vendor’s site.
Work around a box narrower than the documentation
Assume less room than the specification promises. The workspace above caps each reply at 4,096 tokens against a documented ceiling of 128K, displays no context figure at all, and leaves the thinking, web-search and streaming controls unavailable. For practical work that means attachments are the only way to bring large material in, and only PDF files up to 24 MB are accepted. Long documents are therefore the realistic route here and images are not, even though the underlying model is documented as text-in and text-out, which makes this a limitation of the local catalogue rather than a property of the model.
Read the price as two meters, not one
The token rate and the search rate are measured separately and do not combine. Tokens are billed on a prepaid, metered basis in two currencies, with a cheaper rate for input that hits the cache, and batch processing is explicitly not supported for this model. The web-search plugin is billed per call rather than per token, at a rate quoted per thousand calls, and it has to be turned on in the console before it runs at all. Nothing on this page converts into currency: this site bills in points against a catalogue reference price, and the official rates above are there to be read as official rates.
What Xiaomi MiMo V2.5 Pro costs, and which official number belongs to which meter
One official price table covers this model in two currencies, and a second charge sits outside it. The domestic table prices a million tokens at 0.025 yuan for input that hits the cache, 3.00 yuan for input that misses it and 6.00 yuan for output. The international table repeats the same three columns in dollars: 0.0036, 0.435 and 0.87. Both were checked on 9 October 2026, and neither attaches a date, a campaign window or an introductory price to this model. Two further official rules anchor the line. This model shares its row with Xiaomi MiMo V2.6 Pro, so the same rates cover both. And batch processing is not supported here, which the pricing page records by naming this model explicitly alongside two others. The charge outside the token rates is the web-search plugin: it is billed per call rather than per token, at 16 yuan or 5 dollars per thousand calls depending on the table, and it has to be enabled before it does anything. Billing itself is prepaid and metered against actual consumption, and a subscription plan exists alongside it with an allowance that does not mix with the prepaid balance. Nothing above is this site’s checkout price. This site bills in points against a catalogue reference price of 67.5676 points per million input tokens and 67.5676 per million output tokens, and points are not a conversion of either currency.
View points and plans ↗Xiaomi MiMo V2.5 Pro FAQ
When does Xiaomi MiMo V2.5 Pro stop working?
The official deprecation notice gives 21 October 2026 at 10:00 Beijing time, and states that requests using the old model name will receive an error message once that time passes. The same notice sets the same deadline for Xiaomi MiMo V2.5, so the two entries retire together. The vendor says rollout of the change will be announced in advance by text message, email and a website notice, and advises moving to a newer model as soon as possible. Note what the notice does not provide, which is the reason this page is careful about it: it names no successor.
Is there a replacement model I should move to?
Not one that the vendor names. The deprecation table records that there is no system replacement model and that the model will be taken offline directly when the deadline passes. That is a deliberate contrast with the previous retirement round, which is documented on the same page and which did name a replacement together with the date the swap would happen. So the two cases are not the same and should not be read as the same. Anything that depends on this exact model identifier should be treated as having an end date, and the choice of what to move to is left to the reader because it is left to the reader by Xiaomi too.
Can I turn thinking off?
Yes, and the parameter is documented. Deep thinking is governed by thinking.type, which accepts enabled or disabled, and it is enabled by default for this model. Turning it off changes the sampling behaviour as well: while thinking is on, the model refuses a custom temperature and top_p and forces both to 1.0 and 0.95, so any tuning you send in that state is ignored rather than honoured. One implementation detail catches people out: thinking is not a standard OpenAI parameter, so a request constructed with the OpenAI SDK has to carry it inside extra_body. The switch is not exposed in this workspace, so this is a description of the API rather than of the box above.
Is Xiaomi MiMo V2.5 Pro open source?
Yes, and the licence is a permissive one. The model page states that the release is fully open-sourced under a permissive licence, and the repository holds the weights, the tokenizer and a complete model card. That model card declares an MIT licence and carries a published architecture summary: a hidden size of 6,144, seventy layers of which one is dense and the rest are mixtures of experts, ten full-attention layers against sixty sliding-window layers, 128 attention heads with 8 key-value heads, 384 routed experts activating 8 per token, and three multi-token-prediction layers. Deployment guidance for SGLang and vLLM is published with it, which is what makes running it yourself a documented path rather than an inference.
How much context do I get here?
The documented window is one million tokens and the documented maximum output is 128K tokens. This interface displays neither. The catalogue this site reads carries no context or maximum-output field for this model, so the composer falls back to its own 4,096-token request limit and no context figure appears anywhere in the workspace. Any number you see quoted for this page is the documented ceiling rather than a live reading of a session, and the same is true for every model here, because the omission comes from how this site reads its catalogue rather than from anything specific to Xiaomi MiMo V2.5 Pro.
What does the web-search plugin cost?
It is billed separately from tokens, by call rather than by consumption. The pricing page quotes it at 16 yuan per thousand calls on the domestic table and 5 dollars per thousand calls on the international one, and the rate includes both web search and page parsing. It is not part of the token price and it is not enabled by default: the guide states that the plugin has to be turned on in the console before it works at all. The workspace above does not expose a search control for this model, so the plugin is not reachable from this page even if it is enabled on an account.
How does it compare with Claude Opus 4.6?
On the vendor’s own eight-column benchmark table, over the eight rows that carry a value on both sides, Xiaomi MiMo V2.5 Pro leads on two and trails on six. It wins the multi-turn agent benchmark at 72.9 against 72.4 and the terminal benchmark at 68.4 against 65.4, and it trails on a general-agent Elo rating, a task-completion evaluation, a graduate-level science exam, two software-engineering sets and an implementation-ranked coding measure. The scores are the vendor’s readings of its own model, and the vendor makes a broader claim elsewhere that this page does not repeat: an overall statement of superiority would not be supportable from the table, because the columns are not one-sided.
Can I use the official API instead of this page?
Yes, and the two routes are priced on different meters. On the API side the model is reached with a key against the vendor’s base URL, is compatible with both the OpenAI and the Anthropic protocol so a migration is mostly a change of base URL and model name, and is billed per token with a cheaper rate for cached input. On this page the same model runs behind a points balance against a catalogue reference price, capped at 4,096 output tokens per reply, with images not accepted as attachments and the thinking, search and streaming controls unavailable. Two constraints belong to the API rather than to this page and are worth knowing in advance: requests using the old model name will fail after 21 October 2026, and batch processing is not supported for this model.
Sources and model information
Information checked
Start a Xiaomi MiMo V2.5 Pro conversation
The chat box above runs this model, and this page is written to be checked rather than trusted. The specifications, the architecture claims, the case studies and the benchmark table come from the vendor’s product page, the rates come from the pricing table that prints the same three columns in two currencies, the end date comes from the deprecation notice, the thinking rules come from the deep-thinking guide, and the limits of this workspace come from the code that runs it. Where the two sides disagree, the page says so in the sections above instead of quietly writing the official number next to this interface. Where the vendor’s own artefacts disagree with each other, the readings are given separately rather than merged. The three most useful facts here are also the three least repeated elsewhere: that the retirement carries no replacement model, that thinking can be switched off and changes the sampling rules when it is, and that the web-search plugin is billed per call rather than per token.
Start a chat with Xiaomi MiMo V2.5 Pro ↑