Mimo is Xiaomi's family of long-context multimodal models plus the browser, agent, terminal and API products built on them, sold through a metered lane and a coding-only subscription with published open weights.
Used this tool? Rate it
Used this tool? Rate it
Mimo (presented on its own pages as Xiaomi MiMo) is a family of large models and the developer products built on them, published by Xiaomi and described on its home page as a universal smart platform. It is not one application: the same models are reached through a browser chat surface, an agent workspace, a terminal client, a desktop client and a metered API.
Two flagship models anchor it. MiMo-V2.5-Pro is the long-horizon reasoning specialist, and MiMo-V2.5 is the natively multimodal member accepting image, video and audio alongside text. Both carry a one-million-token context window and a maximum output of 128K tokens, which puts repository-scale and long-document work at the centre of gravity rather than short chat turns.
The weights are published rather than held back: Xiaomi's model organisation lists downloadable checkpoints including a one-trillion-parameter Pro release and a 311B-parameter MiMo-V2.5, so hosted and self-hosted routes can be compared directly.
One structural detail shapes everything below: access is sold in two billing lanes with separate credentials and base URLs, and that choice constrains what you are contractually allowed to build.
The API spans text generation, image, audio and video understanding, speech recognition, speech synthesis, deep thinking, structured output, tool calling and web search. Speech sits in dedicated models: MiMo-V2.5-ASR handles recognition across Chinese, English and regional dialects, and the MiMo-V2.5-TTS series covers preset voices, timbre design and timbre cloning.
Capability is not uniform across the family. Image understanding is documented as supported by mimo-v2.5 only, so a pipeline standardised on the Pro model for reasoning quality cannot also send it pictures, and web search is likewise restricted to the two flagship text models.
The speech models carry far tighter envelopes: recognition runs with an 8K context and 2K maximum output, synthesis with 8K context and 8K output, making them request-scoped rather than long-context services.
Ceilings are published per modality, not as one platform rule. A single image supplied by URL must not exceed 50 MB, the same ceiling applies to its base64-encoded string, and pixel counts are bounded between 8,192 and 8,388,608.
Video is more generous by URL and stricter inline: a single video file must not exceed 300 MB by URL, while its base64 string must stay under 50 MB. Containers accepted are MP4, MOV, AVI and WMV, frame sampling uses an fps parameter in the range 0.1 to 10, and extraction stops at 2,048 frames.
Deep thinking runs on both flagship models and changes the request itself, not just answer quality. Those models do not support custom temperature and top_p parameters, enforcing 1.0 and 0.95 instead, so sampling cannot be tuned while the mode is on.
The multi-turn rule matters more. Combined with tool calls, the reasoning field must be passed back completely or the API returns a 400 error, and omitting it degrades instruction following and increases hallucinations. Agent loops stripping intermediate fields to save tokens break here.
Web search runs forced or triggered by the model's own intent recognition, tuned by parameters for keyword breadth, result limits and approximate user location. It is a metered plugin rather than a free toggle: it must be activated before use, and a five-minute cache period applies after enabling or disabling it.
Rate limits are published per model and aggregated per account: both flagships allow 100 requests and 10M tokens per minute, while speech recognition allows 100 requests against a far smaller 10K token budget. The ceiling counts combined traffic from every key on one account calling that model.
Because the endpoints mirror two widely used protocols, the documented adoption path is to reconfigure a tool you already run. Documented integrations include the vendor's own MiMo Code alongside third-party clients such as OpenCode, OpenClaw, Claude Code, Codex, Hermes Agent, Kilo Code, Cline, Cherry Studio, Qwen Code and CodeBuddy, each configured identically: point the base URL at the OpenAI-compatible or Anthropic-compatible endpoint, then supply the key matching your lane.
Repository-scale coding sessions. MiMo Code is a terminal-native assistant that reads and writes code, runs commands, manages Git and keeps a persistent memory system so it retains an understanding of a project across sessions. Agent modes for build, plan and compose work, subagent orchestration with parallel execution and MCP connections suit multi-step refactors rather than single-file edits.
Long-document and mixed-media analysis. The million-token window with native image, video and audio input covers archives that would otherwise need chunking, if the work is routed to the model that accepts those inputs.
Agent workflows that produce deliverables. MiMo Claw is an agent workspace rather than a chat box: it generates, previews, edits and exports documents through an integrated office suite, manages files with batch download and runs multiple conversations and tasks in parallel.
Voice products. Timbre cloning from a single sample, timbre design and preset voices are separate models, so a product picks the control level it needs rather than accepting one fixed voice.
Evaluation before commitment. Published checkpoints let a team benchmark the hosted endpoint against a self-hosted copy of the same weights first.
Good fit: developers already running an OpenAI-compatible or Anthropic-compatible client, since migration is mostly configuration and the documented integration list covers most terminal and editor workflows.
Good fit: teams whose work is genuinely long-context. A million-token window with 128K output only earns its price when documents, repositories or agent trajectories reach that scale.
Mixed fit: teams hoping to build a product backend on a subscription. That lane is contractually limited to coding tools, so backend and automation traffic must sit on the metered lane and a team needing both funds two systems.
Poor fit: buyers needing organisation accounts, refunds or a verifiable privacy commitment before signing. Login is personal-account only, refunds are not currently supported, and the privacy policy and service agreement pages returned no readable text this round, so those commitments could not be confirmed at source.
Poor fit: teams in the EU, the UK or Korea needing the preview models, documented as not currently available in those markets.
Access spreads across distinct surfaces rather than one application:
The terminal client's source is released under the MIT licence, with separate use restrictions and a trademark policy alongside it.
Two lanes sell side by side, and the pay-as-you-go balance is explicitly not interoperable with the subscription quota. Metered rates are published per million tokens, in yuan for China and dollars overseas:
| Model | Region | Cache hit input | Cache miss input | Output |
|---|---|---|---|---|
| mimo-v2.5-pro | Overseas | $0.0036 | $0.435 | $0.87 |
| mimo-v2.5 | Overseas | $0.0028 | $0.14 | $0.28 |
| mimo-v2.5-pro | China | ¥0.025 | ¥3.00 | ¥6.00 |
| mimo-v2.5 | China | ¥0.02 | ¥1.00 | ¥2.00 |
Two line items sit outside that table. Speech recognition bills by audio duration at $0.074 per hour overseas and ¥0.5 per hour in China; web search bills at $5 per thousand requests overseas and ¥16 per thousand in China, on top of the tokens retrieved pages consume. Cache writes and the speech synthesis models are free for a limited time, so today's quoted total is not a stable forecast.
The subscription lane sells four tiers with monthly credit allowances: Lite at $6 a month with 4.1B credits, Standard at $16 with 11B, Pro at $50 with 38B and Max at $100 with 82B, annual purchases costing roughly twelve times the monthly figure. When a monthly allowance is exhausted the system stops service rather than falling back to metered billing, and refunds are not currently supported.
An independent software directory files this product as a large language model tool and names the assistants buyers weigh against it, including ChatGPT, DeepSeek, Claude, Google Gemini, Qwen, Kimi and Microsoft Copilot. The signal there is that the comparison set is the general assistant market, not a niche.
For the terminal client the comparison is narrower and better measured. Independent reporting records MiMo Code at 62% on SWE-bench Pro against 55% for Claude Code paired with Claude Sonnet 4.6, and 73% on Terminal Bench 2 against 69%. The same coverage notes Xiaomi did not publish comparisons against OpenAI's Codex or Google's Gemini CLI, and that on the public leaderboard GPT-5.5 scored 82.7% on Terminal Bench 2, above the figure claimed here.
On the model itself, an independent measurement service scores MiMo-V2.5-Pro at 26 on its intelligence index, well above the median of 17 for open-weight models of similar size, while reporting output at 35.3 tokens per second against a peer median of 66.7. That result is bought with speed.
Access is not locked to one seller. The multimodal model is resold through at least one marketplace at rates and a context ceiling differing from the vendor's own, so a price comparison is worth running before standardising on the first-party endpoint.
Not entirely. Metered API use is charged per token from the first call, while the subscription starts at $6 a month. Some parts are temporarily free: cache writes and the speech synthesis models are listed as free for a limited time, and the agent workspace shipped a free beta before its paid version.
No. The subscription package is restricted to programming tools and prohibits API calls for clearly non-coding scenarios such as automated scripts and custom application backends, with service suspension and key banning as the stated consequence. Backends belong on the metered lane.
Send them to mimo-v2.5. Image understanding is documented for that model rather than the whole family, and a single image is capped at 50 MB with pixel counts bounded between 8,192 and 8,388,608.
Yes, at least in part. Checkpoints are published for download and the terminal client's source is released under the MIT licence. Self-hosting a trillion-parameter model still demands serious infrastructure, so this is an option for evaluation and control rather than a cheap escape.
Independent measurement puts output at 35.3 tokens per second against a peer median of 66.7, and time to first token at 3.40 seconds against a median of 2.32 seconds, while scoring intelligence well above that peer median. The trade is quality for latency.
Service stops. The documentation states that when the monthly total quota of the package is exhausted the system stops service, with no automatic fall-through to metered billing. Refunds are not currently supported, so sizing the tier correctly matters.
