Xiaomi MiMo is worth testing, not blindly switching to. The source-backed case is unusually strong for a newer model lane: official Xiaomi MiMo docs list OpenAI-compatible and Anthropic-compatible APIs, 1M-context V2.5 models, a coding-focused Token Plan, tool setup paths, and open weights on Hugging Face. The caution is just as important: pricing changed in late May 2026, legacy V2 model names are being routed to V2.5 before deprecation, and privacy/enterprise guarantees need direct review before sensitive production use.
Best next click: Smart Spend Guide
Quick Facts
| Item | Current Xiaomi MiMo status |
|---|---|
| Provider | Xiaomi MiMo / Xiaomi |
| Main developer model family | MiMo-V2.5 series |
| Current flagship API model | mimo-v2.5-pro |
| Current multimodal API model | mimo-v2.5 |
| Other current API models | mimo-v2-flash, mimo-v2.5-asr, mimo-v2.5-tts, mimo-v2.5-tts-voiceclone, mimo-v2.5-tts-voicedesign, legacy mimo-v2-pro, mimo-v2-omni, and mimo-v2-tts during migration |
| Context and output limits | V2.5 Pro and V2.5: 1M context and 128K max output in the model/rate-limit docs |
| API formats | OpenAI-compatible chat completions and Anthropic-compatible messages |
| Plan types | Pay-as-you-go API, Token Plan subscription for programming tools, referral/trial credits, and open-weight self-hosting |
| Open-weight status | MiMo-V2.5-Pro and MiMo-V2.5 are listed on Hugging Face with MIT licenses; Xiaomi says the V2.5 series weights are open-sourced |
| Pricing status | Source-backed from Xiaomi docs updated June 2, 2026; checkout remains final authority |
| Biggest unknowns | Enterprise data controls, zero-retention guarantees, regional availability by account, actual tool quota burn, latency, and independent eval results |
Which Xiaomi Path Should I Choose?
| If you want… | Start with | Why | Stop if… |
|---|---|---|---|
| A normal API for apps, scripts, evals, and custom routing | Pay-as-you-go API | It uses ordinary Xiaomi MiMo API keys and account balance, with OpenAI-compatible and Anthropic-compatible routes | You need flat monthly coding-tool quota or cannot accept hosted API data terms |
| A coding-tool subscription lane | Token Plan | It is built for programming tools such as OpenCode, OpenClaw, and Claude Code, with package API keys and dedicated base URLs | You need backend/API automation, non-coding use, refund flexibility, or PAYG balance interoperability |
| A fast hands-on model trial | MiMo Studio / demo paths | It lets you inspect model behavior before wiring API keys into tools | You need reproducible API metering, privacy review, or production terms |
| Local control or fine-tuning | Open weights | V2.5 and V2.5-Pro weights are MIT-licensed on Hugging Face | You do not have server-grade inference hardware and operations capacity |
What Xiaomi Actually Offers
There are two different Xiaomi AI stories here. Do not mix them.
Developer model lane: Xiaomi MiMo is the model/API platform. It includes hosted API access, Token Plan subscriptions for coding tools, open weights, Studio access, model docs, and setup paths for agent/coding tools.
Device and ecosystem lane: Xiaomi also ships AI inside phones, apps, smart-home products, cars, and HyperOS-era experiences. Those may use Xiaomi AI systems, but they are not automatically the same thing as a public MiMo API plan. This guide only treats device/app AI as relevant when Xiaomi’s developer docs connect it to MiMo models, APIs, or open weights.
The current developer shortlist is:
| Model | What it is | Use it for |
|---|---|---|
mimo-v2.5-pro | Flagship sparse MoE language model, 1.02T total / 42B active parameters on the Hugging Face card | Hard coding, long-horizon agent work, long-context reasoning |
mimo-v2.5 | Native omnimodal model, 310B total / 15B active parameters on the Hugging Face card | Text, image, video, audio understanding, multimodal agent workflows |
mimo-v2-flash | Lower-cost V2 flash model still listed in API pricing and model limits | Fast and cheaper text-generation lane |
mimo-v2.5-asr | ASR model | Speech recognition |
mimo-v2.5-tts family | TTS, voice clone, and voice design models | Speech synthesis experiments |
Legacy names matter. Xiaomi’s docs say mimo-v2-pro and mimo-v2-omni auto-route to V2.5 pricing from June 1, 2026, and are scheduled for full deprecation on June 30, 2026. New integrations should use the V2.5 names directly.
Release Chronology
This is the compact source-backed sequence that matters for buyers and implementers:
| Date | Event | Why it matters |
|---|---|---|
| 2025 open-weight lane | MiMo-7B models appeared under the XiaomiMiMo Hugging Face organization | Xiaomi had smaller open-weight research models before the hosted V2 API lane |
| December 16, 2025 | MiMo-V2-Flash release appears in Xiaomi’s model-release log | Establishes the V2 flash/code-agent foundation model before V2-Pro and V2.5 |
| January 12 / February 4, 2026 | MiMo-V2-Flash updates appear in Xiaomi’s model-release log | Shows Xiaomi was iterating the same V2-Flash model ID before the later V2.5 migration |
| March 18, 2026 | MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS releases appear in Xiaomi’s model-release log | These are now legacy names for new integrations because the docs route them toward V2.5 |
| April 22-23, 2026 | MiMo-V2.5 launch/public beta and V2.5-Pro release | V2.5 introduces the current 1M-context multimodal/agent lane and the V2.5-Pro coding/agent lane |
| May 27, 2026 | Xiaomi price update takes effect | PAYG V2.5 prices drop and Token Plan billing/quota mechanics change |
| May 28, 2026 | V2.5 series open-weight announcement update | Xiaomi says V2.5 weights are open-sourced under MIT and usable for commercial inference/fine-tuning without extra authorization |
| June 1, 2026 | mimo-v2-pro and mimo-v2-omni begin auto-routing to V2.5 pricing | Old model names can hide new billing/model behavior |
| June 2, 2026 | Current model/rate-limit and Token Plan docs show updated tables | This is the current page-level anchor for limits, plan quotas, and package rules |
| June 30, 2026 | V2-Pro and V2-Omni deprecation deadline in Xiaomi docs | New work should use mimo-v2.5-pro or mimo-v2.5 directly |
Plans And Pricing
Use the status labels consistently:
| Lane | Status | What Xiaomi documents | Caveat |
|---|---|---|---|
| Open weights | Free | V2.5 model weights are published under MIT on Hugging Face; Xiaomi says commercial inference deployment and secondary training need no extra authorization | Self-hosting a 310B or 1.02T model is not free in infrastructure terms |
| Referral credits | Free | Refer-and-earn gives both users API bonus credits, valid 40 days, API calls only | Not usable for Token Plan packages; daily reward issuance can be limited |
| TTS limited-time API use | Free | Xiaomi lists the V2.5 TTS series as free for a limited time | “Limited time” is not an evergreen free tier |
| Pay-as-you-go API | Paid | Overseas V2.5 pricing lists mimo-v2.5-pro at cache-hit/input/output rates of $0.0036 / $0.435 / $0.87 per 1M tokens, and mimo-v2.5 at $0.0028 / $0.14 / $0.28 per 1M tokens | Pricing changed recently; account region, balance, and checkout/docs control final billing |
| Token Plan | Paid | Monthly plans list Lite, Standard, Pro, and Max at $6, $16, $50, and $100 with fixed credit quotas | Credits are not normal API dollars; quota burn depends on model, cache status, and usage path |
| Xiaomi versus other models | Compare | Xiaomi’s API and Token Plan are now price-competitive enough to test beside GLM, MiniMax, Kimi, Gemini, and local/open-weight lanes | Do not call it a winner without your own evals |
The Token Plan is the most important caveat. Xiaomi describes it as a subscription for AI programming scenarios that works in tools such as OpenCode, OpenClaw, and Claude Code. It uses package-specific API keys and base URLs, and Xiaomi says Token Plan package quota is not interoperable with ordinary pay-as-you-go account balance.
The plan table in Xiaomi’s June 2 docs lists:
| Token Plan | Monthly price | Monthly fixed credit limit | Annual price | Annual fixed credit limit |
|---|---|---|---|---|
| Lite | $6/month | 4.1B credits | $63.36/year | 49.2B credits |
| Standard | $16/month | 11B credits | $168.96/year | 132B credits |
| Pro | $50/month | 38B credits | $528/year | 456B credits |
| Max | $100/month | 82B credits | $1,056/year | 984B credits |
Do not flatten those credits into a single per-token price. Xiaomi’s docs list separate package-credit consumption for cache-hit input, cache-miss input, output, and ASR duration. The safe buying question is: does your actual coding tool show acceptable credit burn on real tasks?
Developer Access Paths
| Access path | Current read |
|---|---|
| API | Xiaomi documents OpenAI-compatible and Anthropic-compatible API endpoints at api.xiaomimimo.com, with API keys created in the console |
| Token Plan tools | Xiaomi documents Token Plan base URLs for China, Singapore, and Europe clusters, with OpenAI-compatible and Anthropic-compatible routes |
| Coding/agent tools | Xiaomi documents setup paths or overview support for OpenCode, Claude Code, OpenClaw, Hermes Agent, Kilo Code, Cherry Studio, Qwen Code, CodeBuddy, Cline, and compatible custom providers |
| Studio/app | Xiaomi links MiMo Studio and “Try it now” paths for interactive use |
| Open weights | XiaomiMiMo publishes V2.5-Pro, V2.5, V2-Flash, ASR, audio, embodied, and vision-language resources on Hugging Face/GitHub |
| Enterprise | Xiaomi references enterprise-level development scenarios and corporate authentication/payment in FAQ material, but this page did not verify a full public enterprise data-control sheet |
For coding tools, the practical split is simple:
- Use ordinary API keys for normal pay-as-you-go API calls and custom applications.
- Use Token Plan keys for Xiaomi’s subscription quota inside supported programming tools.
- Use open weights when you can afford the infrastructure and need local control.
Benchmarks And Capability Claims
Keep the source labels attached. Xiaomi has useful signals, but none of them are an AIHackers-owned “switch now” result.
| Evidence bucket | What it says | How to use it |
|---|---|---|
| Xiaomi-reported launch claims | The MiMo-V2.5 launch page says V2.5 supports native visual/audio understanding, 1M context, and agentic capability. It also makes benchmark-specific relative claims against frontier closed models on video and multimodal agentic work. | Treat as vendor positioning and a reason to test multimodal/agent tasks. Do not convert it into “beats Claude/Gemini” headline language. |
| XiaomiMiMo model-card reported | The V2.5-Pro card reports 1.02T total / 42B active parameters, 1M context, base-model eval tables, long-context GraphWalks results, and deployment guidance. The V2.5 card reports 310B total / 15B active parameters, 1M context, text/image/video/audio modalities, and training on about 48T tokens. | Use for architecture, context, open-weight, and model-card benchmark references. Keep “reported by model card” visible. |
| arXiv lineage | MiMo-7B, MiMo-V2-Flash, and MiMo-VL papers document Xiaomi’s research lineage and earlier model families. | Use for context only. They are not proof of current hosted mimo-v2.5 or mimo-v2.5-pro API behavior, pricing, latency, or coding quality. |
| Third-party / leaderboard signal | Hugging Face surfaces evaluation entries such as SWE Bench Resolved, WildClawBench, and Claw-Eval on the V2.5-Pro page. | Useful shortlist signal, especially for coding-agent work, but still not a substitute for your own repo evals. |
| Not AIHackers-owned | No AIHackers-owned Xiaomi patch-quality benchmark, latency benchmark, privacy test, long-running coding-agent field test, or Token Plan burn study has been run yet. | Keep Xiaomi in the test bench, not the default stack, until real workload checks pass. |
That makes the verdict narrow: Xiaomi MiMo belongs in an eval stack. It does not yet replace Claude, GPT, Gemini, Kimi, GLM, MiniMax, or local models by default.
Privacy, Terms, And Data Use
The hosted platform is not the same risk profile as open weights.
What was verified:
- Xiaomi publishes a MiMo Open Platform privacy policy and service agreement.
- The API requires a Xiaomi account, and the docs currently describe personal account login for the Open Platform.
- FAQ material says Token Plan purchases require personal or corporate real-name authentication.
- Domestic and overseas accounts can receive different base URLs and keys, and Xiaomi says domestic/overseas usage is calculated separately.
What was not verified:
- A clear public zero-retention promise for ordinary API calls.
- A public enterprise data-isolation or no-training term that can be summarized safely without legal review.
- Whether all regions receive the same privacy, payment, and support terms.
For private code, regulated data, client data, or enterprise secrets, use open weights or obtain written terms before routing production data through the hosted Xiaomi API.
Where Xiaomi Fits
| Compared with | Xiaomi is interesting when… | Prefer the alternative when… |
|---|---|---|
| Qwen 3.6 Plus | You want official Xiaomi API access, Token Plan tools, or MiMo open weights instead of a hosted preview lane | You want a free 1M-context OpenRouter/OpenCode eval before paying |
| Kimi k2.5 | You want to test lower listed V2.5 API output rates, 1M context, and Xiaomi’s coding-tool Token Plan | You already rely on Kimi’s official coding lane, membership quota, or Kimi API behavior |
| GLM-5.1 | You want another low-cost supported-tool lane with open weights and multimodal/audio models | You want Z.AI’s established GLM Coding Plan path and current GLM-5.1 guide |
| MiniMax M3 | You want a second 1M-context coding-agent Token Plan candidate and MiMo’s MIT-weight option | You want MiniMax’s current M3 buyer guide and already know its Token Plan fits your tools |
| Gemini 3 Flash | You want a Chinese open-weight/API model lane to test beside Google high-context workflows | You need Google-native AI Studio, Vertex, or Gemini app integration |
| Gemma 4 | You have serious hardware and want to self-host a larger Xiaomi model | You want a cleaner local/on-device lane for laptops, phones, and smaller private deployments |
The simplest Xiaomi role today: add MiMo to your test bench as a low-cost long-context and coding-agent candidate. Keep premium models for review and arbitration until MiMo passes your own evals.
Evaluation Plan
Run the same tasks you use for GLM, MiniMax, Kimi, Gemini, GPT, and Claude:
| Test | What to ask Xiaomi MiMo | Pass signal |
|---|---|---|
| Coding | Fix a real failing test with repository context | Small patch, correct diagnosis, no unrelated churn |
| Long context | Analyze a large codebase or docs set | Specific references, accurate cross-file reasoning |
| Multimodal | Feed a screenshot, chart, or UI artifact | Correct visual read and useful implementation advice |
| Tool use | Use a coding tool route with function/tool calls | Stable tool calls, correct continuation behavior, no runaway loops |
| Latency | Run repeated short and long prompts | Predictable response time for your workflow |
| Quota | Track Token Plan credit burn on real tasks | Credit use matches docs closely enough to budget |
| Privacy | Review terms against your data class | Written policy fit before sensitive production use |
Graduate Xiaomi only after it beats your current low-cost lane on your own workload. If it merely looks good on public benchmarks, keep it experimental.
Related Links
- /value/smart-spend/ - paid-stack strategy and low-cost lane triage
- /models/minimax-m3/ - another 1M-context value coding model to test
- /models/glm-5.1/ - current Z.AI supported-tool coding lane
- /models/kimi-k2.5/ - Kimi model guide
- /models/qwen-3.6-plus/ - free 1M-context preview lane
- /models/gemini-3-flash/ - Google high-context value lane
- /models/gemma-4/ - private local/on-device lane
- /compare/models/budget-tier/ - budget model comparison
Sources
- Xiaomi MiMo home
- Xiaomi MiMo-V2.5 launch page
- Xiaomi MiMo model and rate limits
- Xiaomi MiMo pay-as-you-go API pricing
- Xiaomi MiMo Token Plan subscription instructions
- Xiaomi MiMo AI tools overview
- Xiaomi MiMo first API call
- Xiaomi MiMo OpenAI API compatibility
- Xiaomi MiMo Anthropic API compatibility
- Xiaomi MiMo refer-and-earn rules
- Xiaomi MiMo FAQ
- Xiaomi MiMo model release log
- Xiaomi MiMo-V2.5 open-source announcement
- Xiaomi MiMo-V2.5 price update
- XiaomiMiMo on Hugging Face
- XiaomiMiMo/MiMo-V2.5-Pro model card
- XiaomiMiMo/MiMo-V2.5 model card
- XiaomiMiMo on GitHub
- MiMo-7B paper
- MiMo-V2-Flash paper
- MiMo-VL paper
Last verified: June 11, 2026. Xiaomi MiMo pricing, plan credits, model routing, promotions, and supported tools can change quickly. Treat Xiaomi’s live docs, console, and checkout as final authority before buying or routing sensitive work.