What is Xiaomi MiMo V2.5?
Xiaomi MiMo V2.5 is a sparse mixture-of-experts model with 310 billion parameters in total and 15 billion active, released as open weights under an MIT licence and free to download and run yourself. Its language backbone inherits the hybrid attention design of MiMo-V2-Flash: sliding-window attention and global attention interleaved at a ratio of five to one with a window of 128 tokens, a pattern the model card says cuts KV-cache storage by nearly six times while a learnable attention-sink bias holds long-context quality. A dedicated vision encoder of 729 million parameters and an audio encoder of 261 million are attached through lightweight projectors, and three multi-token-prediction modules of 329 million parameters accelerate inference. Pre-training ran over roughly 48 trillion tokens in FP8 mixed precision, and the context window was extended in stages from 32K to 256K to the full million during post-training. Three relationships are worth fixing before the cards below. First, this is not its larger sibling in a smaller case. The parameter count, the pre-training volume, the attention ratio and the input modality all differ, and the vendor’s model page records the input modality as text, image, video and audio. Second, the retirement is handled rather than abrupt: the notice names MiMo V2.6 Flash as the system replacement model, states that requests are routed to it automatically from 14 October 2026, and publishes the one parameter default that changes as a result. Third, the window that is documented and the window that is displayed here are not the same thing, and the same is true of the price.
What Xiaomi MiMo V2.5 is worth using for
Text, image, video and audio in, and a file picker here that offers PDFs only
This is where this model parts company with most of the text models on this site. The vendor’s model page records the input modality as text, image, video and audio, and the three multimodal guides list this model in their supported set alongside the V2.6 generation. Two details come from those guides and are worth knowing before you plan around them. Video input takes an fps field that defaults to 2 and accepts values from 0.1 to 10, together with a media_resolution setting that defaults to the standard tier, and audio is counted at roughly 6.25 tokens for every second of sound. The workspace above reads a catalogue that declares neither vision nor multimodal input for this model, so its file picker offers PDF files up to 24 MB and nothing else. That is not a correction to the specification, it is the local wiring: the same model can be omni-modal in the documentation and text-and-PDF here, and this page reports both rather than picking one.
Here is a screen recording and a screenshot of the same failure. Watch the recording, read the screenshot, and tell me at which step in the recording the screenshot was taken. If the recording is ambiguous, say which part you cannot resolve.
A published end date with a successor named, and the one default that changes
Retirements are usually announced as a deadline. This one is announced as a handover, and the difference matters for planning. The notice sets 21 October 2026 at 10:00 Beijing time as the moment the old model name stops working, and it also sets 14 October 2026 at 18:00 Beijing time as the moment requests begin routing to MiMo V2.6 Flash, one week earlier. It then does something a bare deadline does not: it publishes the single parameter whose default value changes when the routing takes effect. max_completion_tokens, the ceiling on the length of one reply, defaults to 32768 on this model and to 131072 on its replacement, so a service that never sets the field explicitly will start receiving longer default allowances on the earlier date. The same notice records that a customised parameter is inherited across the switch, and that the replacement models force temperature and top_p to 1.0 and 0.95 in thinking mode. Read all three dates and rules as planning inputs, and re-check the notice before you build anything that depends on this exact model name.
I am deciding whether to keep building on this model name. List the facts that would change that decision, separate the ones that are published from the ones you are inferring, and tell me what I should verify on the vendor’s notice before I commit.
310 billion parameters, five-to-one attention, and the cheapest row in the vendor’s own price table
The model is deliberately the smaller one in its family, and the vendor’s own numbers say what that buys. Total parameters are 310 billion against 1.02 trillion in the larger sibling, active parameters are 15 billion against 42 billion, and the attention pattern runs five to one rather than six, which the model card credits with a KV-cache reduction of nearly six times. Pre-training took roughly 48 trillion tokens, a larger corpus than the larger model was trained on. On price the two sit in different rows of the same table: this model shares a row with MiMo V2.6 Flash at 0.02 yuan for cached input, 1 yuan for uncached input and 2 yuan for output per million tokens, against 0.025, 3 and 6 yuan for the larger one. The vendor also publishes a claim this page repeats only as a claim: that the model matches the larger sibling at half the cost on its internal coding benchmark. Every figure in this card is the vendor’s reading of the vendor’s model.
Compare the cost of running this workload on a large frontier model against running it on a smaller one, using only the figures I give you. Show the arithmetic, state every assumption, and mark which numbers are mine rather than measured.
Thinking that is on by default, can be switched off, and carries a 400 rule
Deep thinking is controlled by a request parameter rather than by a model choice. The field is thinking.type, it accepts enabled or disabled, and it is on by default for this model, which makes switching it off a supported configuration rather than an unsupported one. Turning it off changes the sampling rules in the same move: while thinking is enabled the model refuses a custom temperature and top_p and forces both to 1.0 and 0.95, so a prompt that tunes them is tuning values the model ignores. One rule decides how long sessions have to be run. When an agent carries tool calls across turns, the full reasoning_content from every earlier assistant message must be sent back with the request, and the guide states that omitting it returns a 400 error, adding that a missing reasoning field leaves the context incomplete, with weaker instruction following and more fabrication as the result. Two smaller details sit on the same page: thinking 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 streamed before the answer, a streaming client receives reasoning_content first and the answer afterwards. None of this is reachable from this page, because the catalogue it reads publishes no reasoning field and the switch above is therefore present and unavailable.
Answer without showing intermediate reasoning first, then redo the same question with the reasoning shown step by step. Keep both answers in this conversation so I can compare them, and tell me where the two disagree.
Official Xiaomi MiMo V2.5 specifications
- Identifier
- mimo-v2.5
- Parameters
- 310 billion total, 15 billion active — sparse mixture of experts
- Architecture
- Sliding-window and global attention interleaved 5:1 with a 128-token window, plus multi-token prediction
- Context window
- 1M tokens
- Maximum output
- 128K tokens; API default 32768, accepted range 1 to 131072
- Input modality
- Text, image, video and audio in; text out
- 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
- About 48 trillion tokens in FP8 mixed precision, context extended 32K to 256K to 1M
- Licence
- MIT — weights, tokenizer and model card on Hugging Face
- API protocols
- OpenAI and Anthropic compatible; rate limit 100 requests and 10 million tokens per minute
Xiaomi MiMo V2.5 and Claude Opus 4.6 on the vendor’s own charts
Every value below is read from one page — the Xiaomi MiMo V2.5 product page — which prints two benchmark charts, one for coding and agentic work and one for multimodal perception, each with its own set of comparison models. Seven rows are reproduced here, and every one of them carries a value on both sides on that page. The reason the set is seven rather than the full twelve is that a single rival has to appear in both charts for the pairing to be one comparison rather than two lists placed next to each other, and Claude Opus 4.6 is the only competitor that does. The rows that go against the model are kept rather than filtered out, and they are the majority: of the seven, Xiaomi MiMo V2.5 leads on three and trails on four. It wins the terminal benchmark, chart question answering and the multimodal reasoning exam; it trails on the internal coding benchmark, the daily-agent text evaluation, software-engineering benchmark work and the multimodal agentic set. These are the vendor’s own readings of its own model against a competitor, not an independent measurement, and no row here predicts the quality of your conversation.
| Official benchmark | Xiaomi MiMo V2.5 | Claude Opus 4.6 |
|---|---|---|
| MiMo Coding Bench · Coding | 71.8 | 77.1 |
| Claw-Eval Text · Agentic | 62.3 | 70.8 |
| Terminal-Bench 2.0 · Coding | 65.8 | 65.4 |
| SWE-Bench Pro · Coding | 56.1 | 57.3 |
| CharXiv RQ · Image Understanding | 81.0 | 77.4 |
| MMMU-Pro · Multimodal | 77.9 | 73.9 |
| Claw-Eval Multimodal · Multimodal Agent | 23.8 | 24.8 |
How to use Xiaomi MiMo V2.5 online
Start from the two dates, and from what changes on the earlier one
The most consequential facts about this model are not in its specification. The official notice sets 21 October 2026 at 10:00 Beijing time as the moment the name stops working, and it sets 14 October 2026 at 18:00 Beijing time as the moment requests begin routing to MiMo V2.6 Flash. It also publishes the parameter whose default changes at that switch, max_completion_tokens, from 32768 to 131072. If your code never sets that field, the routing date is the one that changes your behaviour, not the deadline. Treat all three readings as planning inputs rather than warnings, and check the notice again before anything goes live 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 model specification page says what the API supports, and its input-modality line is the one the cards above quote. The pricing page says what a token costs, which features are excluded, and how the search plugin is billed. The product page is where the architecture claims and the benchmark charts 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 quietly picking one, and where the vendor’s own two pages disagree with each other, both readings are given separately.
Plan for 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, search and streaming controls unavailable. The consequence that catches people out here is attachments: the model is documented as reading images, video and audio, and this page accepts PDF files up to 24 MB and nothing else. Long documents are therefore the realistic route here and images are not, which is a limitation of the local catalogue rather than a property of the model.
Read the price as two meters, and note the row it shares
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 this model shares its price row with MiMo V2.6 Flash, which means the same rates cover both. Batch processing is explicitly not supported for this model, which the pricing page records by naming it alongside two others. 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 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 0.02 yuan for input that hits the cache, 1.00 yuan for input that misses it and 2.00 yuan for output. The international table repeats the same three columns in dollars: 0.0028, 0.14 and 0.28. Both were checked on 10 October 2026, and neither attaches a date, a campaign window or an introductory price to this model. Three further official rules anchor the line. This model shares its row with Xiaomi MiMo V2.6 Flash, so the same rates cover both. Batch processing is not supported here, which the pricing page records by naming this model explicitly alongside two others. And the rate is the same on the subscription side as well, where a million tokens costs 2 credits on cache hits, 100 on cache misses and 200 on output. 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 24.3243 points per million input tokens and 48.6486 per million output tokens, and points are not a conversion of either currency.
View points and plans ↗Xiaomi MiMo V2.5 FAQ
When does Xiaomi MiMo V2.5 stop answering?
The official deprecation notice gives 21 October 2026 at 10:00 Beijing time, and states that requests using the old model name receive an error message once that time passes. The same notice sets the same deadline for Xiaomi MiMo V2.5 Pro, so the two entries retire together. The vendor says 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 this notice does not leave open, which is the reason this page is more specific than a deadline alone: it names the successor model and the date on which traffic starts moving to it.
Is there a replacement model I should move to?
Yes, and the vendor names it rather than leaving the choice open. The deprecation table records MiMo V2.6 Flash as the system replacement model for this identifier, with automatic routing beginning at 18:00 Beijing time on 14 October 2026. That matters beyond the routing itself, because the destination is a different generation of model in the same family: the notice describes the change as one of parameter defaults rather than of interfaces, and states that a parameter customised on this model is inherited across the switch. Anything that depends on this exact identifier should be treated as having an end date, and the replacement path is published rather than implied.
What changes when my requests are routed to MiMo V2.6 Flash?
The notice publishes one behavioural change and one parameter change, and both are specific. The parameter change is max_completion_tokens, whose default value is 32768 on this model and 131072 on its replacement, so a service that never sets the field explicitly will start receiving longer default allowances after the routing date. The behavioural change concerns sampling: the replacement models do not accept a custom temperature or top_p in thinking mode and force both to 1.0 and 0.95. The notice also states that parameters you customised on this model are inherited when traffic moves, which means an explicit setting travels with the request rather than being reset.
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, which makes disabling it a supported configuration rather than an unsupported one. 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.
What kinds of input does it accept?
The vendor’s model page records four: text, image, video and audio, with text as the output. Three multimodal guides list this model in their supported set, and they add the details worth knowing before you plan around them. Video input carries an fps field that defaults to 2 and accepts values from 0.1 to 10, plus a media_resolution setting that defaults to the standard tier, and audio is counted at roughly 6.25 tokens per second of sound. Neither guide supports uploading a local file directly, so both route through a hosted URL. This page is the narrower case: the catalogue it reads declares no vision or multimodal capability for this model, so attachments here accept PDF files up to 24 MB and nothing else.
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.
Is Xiaomi MiMo V2.5 open source?
Yes, and the licence is a permissive one. The product page states that the series is fully open-sourced, and the repository holds the weights, the tokenizer and a complete model card. That model card declares an MIT licence and publishes an architecture table in full: a hidden size of 4096, forty-eight layers of which one is dense and the rest are mixtures of experts, nine full-attention layers against thirty-nine sliding-window layers, sixty-four attention heads with eight key-value heads on the global layers and four on the sliding-window ones, 256 routed experts activating eight per token, and three multi-token-prediction layers. A second, shorter-context variant named MiMo-V2.5-Base is published alongside it at 256K. Deployment guidance for SGLang and vLLM is published with the card, which is what makes running it yourself a documented path rather than an inference.
How does it compare with Claude Opus 4.6?
On the vendor’s own product page, which prints two benchmark charts with their own comparison sets, Claude Opus 4.6 is the only rival that carries a value in both. Across the seven rows where both columns are filled, Xiaomi MiMo V2.5 leads on three and trails on four. It wins the terminal benchmark at 65.8 against 65.4, chart question answering at 81.0 against 77.4 and the multimodal reasoning exam at 77.9 against 73.9; it trails on the internal coding benchmark, the daily-agent text evaluation, software-engineering benchmark work and the multimodal agentic set. 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 these charts, because the rows are not one-sided.
Sources and model information
Information checked
Start a Xiaomi MiMo V2.5 conversation
The chat box above runs this model, and this page is written to be checked rather than trusted. The specifications, the architecture claims and the benchmark charts come from the vendor’s product page, the input modality comes from the model specification page, the rates come from the pricing table that prints the same three columns in two currencies, the end date and the replacement model come 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 model reads four modalities while this page accepts one file type, that the retirement names its successor and the single default that changes with it, and that the same model sits in the cheapest row of the vendor’s own price table.
Start a chat with Xiaomi MiMo V2.5 ↑