---
title: "MiMo (Xiaomi)"
source: "https://systems-analysis.info/eng/MiMo_(Xiaomi)"
wiki: "systems-analysis.info/eng"
article: "MiMo_(Xiaomi)"
language: "en"
categories:
  - "Category:English"
  - "Category:Large language models"
  - "Category:LLM families"
  - "Category:Machine learning"
  - "Category:Technology"
revision_id: 249
wiki_created_at: 2026-09-06T22:19:30Z
wiki_modified_at: 2026-09-06T22:19:30Z
downloaded_at: 2026-09-07T22:22:10Z
---

# MiMo (Xiaomi)

**MiMo**, also styled **Xiaomi MiMo**, is a family of [large language models](https://systems-analysis.info/eng/Large_language_model "Large language model"), multimodal foundation models, speech and audio models, embodied-AI models, and agent-oriented systems developed by Xiaomi. The models are principally associated with the Xiaomi MiMo and LLM-Core teams. Xiaomi describes MiMo as its “universal smart platform” for language, multimodal perception, voice, software engineering, agents, and interaction with physical or digital environments.<sup>[\[1\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoPortal-1)</sup>

Official materials do not define *MiMo* as a fixed technical acronym. The name is used for several architecturally distinct branches:

- compact dense reasoning models such as **MiMo-7B**;
- vision-language systems such as **MiMo-VL-7B**;
- sparse [Mixture-of-Experts](https://systems-analysis.info/eng/Mixture-of-Experts_(MoE) "Mixture-of-Experts (MoE)") (MoE) language and agent models in the **MiMo-V2** and **MiMo-V2.5** generations;
- speech-understanding and speech-generation systems, including **MiMo-Audio**, **MiMo-V2.5-ASR**, and the **MiMo-V2.5-TTS** series;
- embodied and graphical-interface models such as **MiMo-Embodied**;
- product systems including MiMo Code, MiMo Claw, MiMo Studio, and the MiMo API.

As of 10 July 2026, **MiMo-V2.5-Pro** is Xiaomi's most capable publicly announced general-purpose language and agent model. It is a sparse MoE model with approximately 1.02 trillion total parameters and 42 billion parameters activated for each token. It uses hybrid sliding-window and global attention, supports a context of up to 1,048,576 tokens, and is intended for long-horizon software engineering, tool use, and agent trajectories containing hundreds or thousands of actions.<sup>[\[2\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Pro-2)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

**MiMo-V2.5** is the principal native multimodal model in the same generation. It contains approximately 310 billion total parameters and activates 15 billion parameters per token. The model accepts text, images, video, and audio, produces text and structured tool calls, and supports a one-million-token context. Its language backbone is supplemented by a 729-million-parameter visual encoder and a 261-million-parameter audio encoder.<sup>[\[4\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25-4)[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)</sup>

The V2.5 product family also includes:

- **MiMo-V2.5-TTS** — speech synthesis using curated voices;
- **MiMo-V2.5-TTS-VoiceDesign** — generation of a new voice from a natural-language description;
- **MiMo-V2.5-TTS-VoiceClone** — voice reproduction from short reference recordings;
- **MiMo-V2.5-ASR** — open-weight automatic speech recognition for Mandarin, English, Chinese dialects, code-switched speech, songs, noisy recordings, and multi-speaker conversations.<sup>[\[6\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25TTS-6)[\[7\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ASR-7)</sup>

MiMo-V2.5-Pro, MiMo-V2.5, MiMo-V2-Flash, MiMo-7B, MiMo-VL, and MiMo-Audio have been distributed with downloadable weights. Several of these checkpoints use the MIT License, while MiMo-V2.5-ASR uses Apache License 2.0. Hosted speech models and product services may use separate proprietary terms. The term *open-weight* is therefore more precise for the family as a whole than implying that every production model, training dataset, and distributed-training pipeline is fully open source.

The original hosted MiMo-V2 endpoints were retired on 30 June 2026, and Xiaomi instructed API customers to migrate to the V2.5 identifiers. The downloadable historical checkpoints remained separate from this hosted-endpoint retirement.<sup>[\[8\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-APIModels-8)[\[9\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-ModelUpdates-9)</sup>

## History and institutional context

### Formation of the MiMo programme

Xiaomi's MiMo programme combines model research, training infrastructure, agent systems, and product integration. The official research organization includes work on pre-training, post-training, multimodality, audio and speech, AI infrastructure, and knowledge systems.<sup>[\[1\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoPortal-1)</sup>

The programme differs from a single-product chatbot effort. Its research outputs span:

- language-model architecture and reasoning;
- multimodal perception;
- reinforcement learning;
- agent environments;
- speech recognition and synthesis;
- embodied intelligence;
- efficient inference;
- sparse attention and long-context systems.

The first major public technical release under the MiMo name appeared in May 2025 with the MiMo-7B family.

### MiMo-7B

The MiMo-7B technical report was published on 12 May 2025. It described a dense decoder-only language model trained from scratch for reasoning tasks.<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

The family included:

- **MiMo-7B-Base**;
- **MiMo-7B-SFT**;
- **MiMo-7B-RL-Zero**, trained through reinforcement learning directly from the base checkpoint;
- **MiMo-7B-RL**, trained through reinforcement learning after supervised fine-tuning;
- later updated reasoning checkpoints with longer generation budgets.

MiMo-7B-Base was trained on approximately 25 trillion tokens. Its development emphasized the density of reasoning patterns in the pre-training corpus rather than treating reasoning only as a post-training capability. The model also used Multi-Token Prediction during pre-training.<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

For reinforcement learning, Xiaomi assembled approximately 130,000 rule-verifiable mathematics and programming problems. The report described a modified GRPO-style optimization process, difficulty-aware code rewards, dynamic sample selection, and a distributed Seamless Rollout Engine.

### MiMo-VL

The **MiMo-VL-7B** vision-language family was introduced in June 2025. It combines:

- a native-resolution Vision Transformer;
- an MLP projector;
- the MiMo-7B language backbone.

Two principal checkpoints were released:

- **MiMo-VL-7B-SFT**;
- **MiMo-VL-7B-RL**.

The training process consumed approximately 2.4 trillion tokens over four stages: projector warm-up, vision-language alignment, broad multimodal pre-training, and long-context supervised fine-tuning. Post-training used **Mixed On-Policy Reinforcement Learning** (MORL), combining rule-verifiable rewards with learned human-preference rewards.<sup>[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)</sup>

MiMo-VL was trained for image and video understanding, visual reasoning, OCR, document analysis, graphical-interface grounding, and multimodal agents.

### MiMo-Audio

The MiMo-Audio research line appeared in late 2025, with an arXiv version dated 29 December 2025 and public model weights subsequently released by Xiaomi.<sup>[\[12\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoAudio-12)</sup>

The released components include:

- **MiMo-Audio-Tokenizer**;
- **MiMo-Audio-7B-Base**;
- **MiMo-Audio-7B-Instruct**.

MiMo-Audio was trained on more than 100 million hours of audio and was designed to unify audio understanding and generation. Its task range includes speech recognition, environmental-audio analysis, speech continuation, voice conversion, style transfer, speech editing, text-to-speech, and spoken dialogue.

The architecture uses a speech tokenizer with residual vector quantization, an audio patch encoder, a language-model backbone, and a patch decoder that reconstructs multi-codebook audio tokens.

### MiMo-Embodied

**MiMo-Embodied-7B** was introduced in November 2025 as a cross-embodied foundation model spanning autonomous driving, robotics, visual navigation, spatial understanding, and task planning.<sup>[\[13\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoEmbodied-13)</sup>

Its developers sought to train one model across datasets that are normally separated by embodiment:

- vehicles;
- mobile robots;
- manipulation systems;
- graphical agents;
- perception and planning environments.

The model was evaluated on embodied-AI and autonomous-driving benchmarks. These results were developer-reported and depended on the supplied action spaces and evaluation scaffolds.

### MiMo-V2-Flash

**MiMo-V2-Flash** was released on 16 December 2025. It was Xiaomi's first large sparse MoE model published under the MiMo name.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

The model contains:

- approximately 309 billion total parameters;
- approximately 15 billion active parameters;
- 48 Transformer layers;
- 256 routed experts per MoE layer;
- eight experts selected per token;
- a hybrid ratio of five sliding-window-attention layers to one global-attention layer;
- a 128-token sliding window;
- three Multi-Token Prediction modules in the post-trained checkpoint;
- a context of up to 256,000 tokens.

MiMo-V2-Flash was trained on approximately 27 trillion tokens. Xiaomi positioned it as an efficient reasoning, coding, and agent foundation model.

### MiMo-V2-Pro, MiMo-V2-Omni, and MiMo-V2-TTS

On 18 March 2026, Xiaomi announced three additional V2 branches.

**MiMo-V2-Pro** scaled the language-and-agent architecture beyond one trillion total parameters, with approximately 42 billion active parameters and a one-million-token context. It retained hybrid attention but increased the local-to-global attention ratio relative to V2-Flash.<sup>[\[15\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Pro-15)</sup>

**MiMo-V2-Omni** added dedicated image, video, and audio encoders to a shared reasoning backbone. It supported structured tool calls, function execution, graphical-interface grounding, and long-form audio and video understanding.<sup>[\[16\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Omni-16)</sup>

**MiMo-V2-TTS** was a large-scale prompted speech-synthesis model using a proprietary audio tokenizer and multi-codebook speech–text modelling. Xiaomi reported pre-training on more than 100 million hours of speech data and further reinforcement learning for prosody, articulation, voice cloning, and expressive control.<sup>[\[17\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2TTS-17)</sup>

These V2 hosted services were later superseded by the V2.5 family.

### MiMo-V2.5

MiMo-V2.5 was announced on 22 April 2026 and added to Xiaomi's API platform during the same release period.<sup>[\[4\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25-4)[\[9\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-ModelUpdates-9)</sup>

The model combines:

- a 310B-A15B language backbone;
- a 729-million-parameter visual encoder;
- a 261-million-parameter audio encoder;
- three Multi-Token Prediction layers;
- a one-million-token context;
- native text, image, video, and audio understanding;
- tool use and agentic post-training.

The complete training programme contained approximately 48 trillion tokens. Xiaomi describes the model as natively omni-modal because multimodal data were incorporated into the main training process rather than only through an inference-time adapter.

### MiMo-V2.5-Pro

MiMo-V2.5-Pro was announced and open-weighted in late April 2026. It enlarged the main language architecture to approximately 1.02 trillion parameters while retaining an active footprint of 42 billion parameters.<sup>[\[2\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Pro-2)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

Compared with MiMo-V2-Pro, the V2.5-Pro release emphasized:

- improved adherence to instructions embedded deep in long contexts;
- longer stable tool trajectories;
- stronger repository-level software engineering;
- lower token use for selected agent tasks;
- greater compatibility with coding and agent scaffolds;
- stronger long-context graph reasoning.

The official release showed developer-selected case studies involving hundreds or thousands of tool calls, including compilation, desktop-application development, and electronic-design automation.

### MiMo-V2.5-Pro-UltraSpeed

On 8 June 2026, Xiaomi and TileRT announced **MiMo-V2.5-Pro-UltraSpeed**, a serving configuration targeting more than 1,000 generated tokens per second on one eight-GPU node.<sup>[\[18\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-UltraSpeed-18)</sup>

UltraSpeed is not documented as an independently pretrained foundation model. It combines:

- selective MXFP4 quantization of MoE experts;
- higher precision for more sensitive non-expert modules;
- DFlash block-masked speculative decoding;
- persistent TileRT execution kernels;
- warp-specialized and heterogeneous GPU pipelines.

Xiaomi released the associated **MiMo-V2.5-Pro-FP4-DFlash** checkpoint. The original public trial was limited and application-based; the API platform continued to advertise restricted UltraSpeed access after the initial period.

## Release chronology

| Date                      | Model or system                         | Publicly disclosed scale                                           | Modalities and role                                                              | Context and distribution                                                    |
|---------------------------|-----------------------------------------|--------------------------------------------------------------------|----------------------------------------------------------------------------------|-----------------------------------------------------------------------------|
| **12 May 2025**           | **MiMo-7B**                             | Approximately 7B dense                                             | Text reasoning, mathematics, code, and general dialogue                          | 32K-class released checkpoints; MIT-licensed open weights                   |
| **4 June 2025**           | **MiMo-VL-7B**                          | Approximately 7B language backbone plus visual encoder             | Text, image, video, OCR, GUI grounding, and visual reasoning                     | Open-weight SFT and RL checkpoints                                          |
| **Late 2025**             | **MiMo-Audio**                          | 7B-class language backbone plus approximately 1.2B audio tokenizer | Audio understanding, speech generation, recognition, conversion, and editing     | Open weights; MiMo-Audio-7B-Base and Instruct                               |
| **20 November 2025**      | **MiMo-Embodied-7B**                    | Approximately 7B                                                   | Autonomous driving, robotics, task planning, and embodied reasoning              | Research checkpoint and technical report                                    |
| **16 December 2025**      | **MiMo-V2-Flash**                       | 309B total / 15B active                                            | Text reasoning, coding, tools, and agents                                        | 256K; open weights under MIT License                                        |
| **January–February 2026** | **Updated MiMo-V2-Flash service**       | Same architectural class                                           | Improved reasoning, coding, instruction following, and tool calls                | Hosted identifier retained while server-side checkpoint was updated         |
| **18 March 2026**         | **MiMo-V2-Pro**                         | More than 1T total / 42B active                                    | Text reasoning, coding, and long-horizon agents                                  | 1M hosted context; proprietary hosted release                               |
| **18 March 2026**         | **MiMo-V2-Omni**                        | Not disclosed                                                      | Text, image, video, audio, tools, and interface grounding                        | 256K-class hosted service                                                   |
| **18 March 2026**         | **MiMo-V2-TTS**                         | Not disclosed                                                      | Prompted speech generation, singing, and voice cloning                           | Hosted speech service                                                       |
| **22–23 April 2026**      | **MiMo-V2.5**                           | 310B total / 15B active                                            | Native text, image, video, and audio understanding; tools and agents             | 1M post-trained checkpoint; MIT-licensed open weights                       |
| **April 2026**            | **MiMo-V2.5-TTS series**                | Not disclosed                                                      | Stock voices, voice design, and voice cloning                                    | Hosted API and MiMo Studio                                                  |
| **April–May 2026**        | **MiMo-V2.5-ASR**                       | Approximately 8B in the published model package                    | Mandarin, English, dialects, code switching, songs, noise, and multi-speaker ASR | Open weights under Apache License 2.0                                       |
| **27 April 2026**         | **MiMo-V2.5-Pro**                       | 1.02T total / 42B active                                           | Text reasoning, coding, software engineering, tools, and long-horizon agents     | 1M post-trained checkpoint; MIT-licensed open weights                       |
| **8 June 2026**           | **MiMo-V2.5-Pro-UltraSpeed**            | Same 1.02T foundation model; selectively quantized                 | High-throughput coding and agent inference                                       | Application-limited service; FP4-DFlash weights released                    |
| **10 June 2026**          | **MiMo Code**                           | Agent product using MiMo models                                    | Long-horizon repository work, terminal tools, memory, and coding workflows       | Hosted product rather than a separate disclosed checkpoint                  |
| **29 June 2026**          | **MOPD publication**                    | Post-training methodology                                          | Multi-teacher capability integration through on-policy distillation              | Research paper; method used in V2 and V2.5 models                           |
| **30 June 2026**          | **Retirement of hosted V2 identifiers** | —                                                                  | Migration from V2 hosted services to V2.5                                        | Applied to API identifiers, not necessarily downloadable historical weights |

<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)[\[12\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoAudio-12)[\[13\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoEmbodied-13)[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)[\[15\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Pro-15)[\[16\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Omni-16)[\[17\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2TTS-17)[\[4\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25-4)[\[2\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Pro-2)[\[18\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-UltraSpeed-18)[\[9\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-ModelUpdates-9)</sup>

## Naming and scope

Several distinctions are necessary when describing MiMo.

- **MiMo is a model ecosystem, not one architecture.** The family includes dense language models, sparse MoE models, visual encoders, audio tokenizers, speech systems, and embodied models.
- **“V2.5-Pro” and “V2.5” are different model scales.** Pro has 1.02T total and 42B active parameters; the native multimodal V2.5 model has 310B total and 15B active parameters.
- **MiMo-V2.5-Pro is text-centred.** It accepts text and tool state and produces text or tool calls. Native image, video, and audio understanding is provided by MiMo-V2.5.
- **Base and post-trained checkpoints have different context limits.** The V2.5 and V2.5-Pro base checkpoints are documented at 256K, while the post-trained versions support up to one million tokens.
- **UltraSpeed is a serving derivative.** It uses quantization and speculative decoding but is not documented as a new general-intelligence checkpoint.
- **MiMo Code and MiMo Claw are products.** Their performance depends on the model, prompts, tool schemas, execution environment, memory, permissions, and product-level orchestration.
- **A “high-speed” or server-side model name is not automatically a new architecture.**
- **Hosted retirement does not delete open weights.** Removal of a model identifier from Xiaomi's API platform is separate from availability of a historical checkpoint on Hugging Face or GitHub.

## Technical foundations

### Autoregressive language modelling

MiMo language models generate text autoregressively. For a token sequence $x_{1},\ldots,x_{T}$:

$$
P_{\theta}(x_{1},\ldots,x_{T}) = \prod\limits_{t = 1}^{T}P_{\theta}\left( x_{t} \mid x_{1},\ldots,x_{t - 1} \right),
$$

where $\theta$ denotes the model parameters.

For a multimodal model, generation can additionally be conditioned on representations of images, video, audio, documents, or tool results:

$$
P_{\theta}\left( y_{t} \mid y_{< t},\mathbf{H}_{text},\mathbf{H}_{vision},\mathbf{H}_{audio},\mathbf{H}_{tools} \right).
$$

This factorization describes the generation interface but does not imply that all MiMo models use the same tokenizer, encoder stack, context length, or post-training system.

### Dense MiMo-7B architecture

MiMo-7B is a conventional dense decoder-only Transformer rather than a sparse MoE model. Its public configurations use approximately:

| Property                            | MiMo-7B                           |
|-------------------------------------|-----------------------------------|
| Transformer layers                  | 36                                |
| Hidden-state dimension              | 4,096                             |
| Feed-forward intermediate dimension | 11,008                            |
| Query heads                         | 32                                |
| Key–value heads                     | 8                                 |
| Positional encoding                 | Rotary position embeddings        |
| Main training context               | 8K, extended to approximately 32K |
| Pre-training tokens                 | Approximately 25T                 |
| Additional prediction objective     | Multi-Token Prediction            |

<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

The use of eight key–value heads with 32 query heads is a form of Grouped-Query Attention, which reduces the key–value-cache footprint relative to conventional multi-head attention.<sup>[\[19\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-GQA-19)</sup>

### Mixture-of-Experts architecture

Beginning with MiMo-V2-Flash, the main language-model line uses sparse expert feed-forward layers.

For token representation $\mathbf{h}$, a simplified MoE transformation is:

$$
\operatorname{MoE}(\mathbf{h}) = \sum\limits_{i \in \mathcal{S}(\mathbf{h})}g_{i}(\mathbf{h})E_{i}(\mathbf{h}),
$$

where:

- $E_{i}$ is an expert feed-forward network;
- $\mathcal{S}(\mathbf{h})$ is the subset selected by the router;
- $g_{i}(\mathbf{h})$ is the corresponding routing weight.

A simplified selection rule is:

$$
\mathcal{S}(\mathbf{h}) = \operatorname{TopK}\left( W_{g}\mathbf{h} \right),
$$

where $W_{g}$ is the learned routing projection.

MiMo-V2.5 contains 256 routed experts and selects eight for each token. MiMo-V2.5-Pro contains 384 routed experts and also selects eight per token.<sup>[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

Sparse activation reduces expert arithmetic. It does not reduce weight storage to the active-parameter count: all experts must remain stored, distributed, or accessible to the serving system.

### Hybrid sliding-window and global attention

Full causal attention has quadratic prefill complexity with sequence length. Standard scaled dot-product attention is:

$$
\operatorname{Attention}(\mathbf{Q},\mathbf{K},\mathbf{V}) = \operatorname{softmax}\left( \frac{\mathbf{Q}\mathbf{K}^{\mathsf{T}}}{\sqrt{d_{k}}} \right)\mathbf{V}.
$$

MiMo-V2 models alternate:

- **Sliding Window Attention (SWA)** — each query attends principally to a limited local region;
- **Global Attention (GA)** — queries can attend across the complete available prefix.

For a local window $w$, a simplified SWA operation is:

$$
\mathbf{o}_{t} = \operatorname{softmax}\left( \frac{\mathbf{q}_{t}\mathbf{K}_{t - w:t}^{\mathsf{T}}}{\sqrt{d_{k}}} \right)\mathbf{V}_{t - w:t}.
$$

MiMo-V2-Flash and MiMo-V2.5 use a five-to-one ratio of SWA layers to global-attention layers with a 128-token window. MiMo-V2.5-Pro uses a six-to-one ratio and the same window size.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

The purpose is to reduce:

- key–value-cache storage;
- long-context memory movement;
- attention computation;
- prefill and decode latency.

Periodic global layers maintain a route for long-range information.

### Attention-sink bias

Aggressive sliding-window attention can degrade quality because tokens may be forced to assign probability mass to unhelpful local positions. MiMo-V2 models use a learnable attention-sink bias, allowing an attention head to allocate little or no probability to the visible tokens when they are not useful.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

Conceptually, an additional learned sink logit $s_{h}$ can be included in the normalization:

$$
\alpha_{t,j} = \frac{\exp(q_{t}k_{j}^{\mathsf{T}})}{\exp(s_{h}) + \sum\limits_{u \in \mathcal{W}_{t}}\exp(q_{t}k_{u}^{\mathsf{T}})}.
$$

The sink is not a normal content token. It provides an alternative destination for attention probability and can reduce distortions caused by a very small local window.

Xiaomi's architecture ablations reported that adding the sink bias restored or improved performance that had been lost by a 128-token sliding window.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

### MiMo-V2.5-Pro architecture

| Property                           | MiMo-V2.5-Pro            |
|------------------------------------|--------------------------|
| Total parameters                   | Approximately 1.02T      |
| Activated parameters per token     | Approximately 42B        |
| Layers                             | 70: one dense and 69 MoE |
| Hidden-state dimension             | 6,144                    |
| Attention heads                    | 128                      |
| Key–value heads                    | 8                        |
| Global-attention layers            | 10                       |
| Sliding-window-attention layers    | 60                       |
| Sliding-window size                | 128 tokens               |
| Query/key head dimension           | 192                      |
| Value-head dimension               | 128                      |
| Routed experts                     | 384                      |
| Experts selected per token         | 8                        |
| Expert intermediate dimension      | 2,048                    |
| Dense-layer intermediate dimension | 16,384                   |
| Multi-Token Prediction layers      | 3                        |
| Post-trained context               | Up to 1,048,576 tokens   |

<sup>[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

### MiMo-V2.5 language backbone

| Property                           | MiMo-V2.5                       |
|------------------------------------|---------------------------------|
| Total parameters                   | Approximately 310B              |
| Activated parameters per token     | Approximately 15B               |
| Layers                             | 48: one dense and 47 MoE        |
| Hidden-state dimension             | 4,096                           |
| Query heads                        | 64                              |
| Key–value heads                    | 8 in global attention; 4 in SWA |
| Global-attention layers            | 9                               |
| Sliding-window-attention layers    | 39                              |
| Routed experts                     | 256                             |
| Experts selected per token         | 8                               |
| Expert intermediate dimension      | 2,048                           |
| Dense-layer intermediate dimension | 16,384                          |
| Multi-Token Prediction layers      | 3                               |
| Post-trained context               | Up to 1,048,576 tokens          |

<sup>[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)</sup>

The published MiMo-V2-Flash technical report used a different key–value-head allocation for its earlier checkpoint: eight KV heads in SWA and four in global attention. Specifications must therefore be attached to the exact checkpoint rather than generalized across the complete V2 family.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

### Multi-Token Prediction

Multi-Token Prediction adds auxiliary modules that predict tokens farther into the future than the ordinary next-token head.

A simplified training objective is:

$$
\mathcal{L}_{MTP} = \mathcal{L}_{1} + \sum\limits_{k = 2}^{K}\lambda_{k}\mathcal{L}_{k},
$$

where $\mathcal{L}_{1}$ is the next-token loss and $\mathcal{L}_{k}$ predicts the token $k$ positions ahead.

At inference time, the MTP modules can serve as a speculative draft system:

1.  several candidate future tokens are proposed;
2.  the main model verifies the candidates;
3.  a correct prefix is accepted in one verification step;
4.  generation resumes from the first rejected position.

MiMo-V2-Flash used one MTP layer during much of pre-training and distributed three adapted MTP modules with the post-trained checkpoint. Xiaomi reported an average accepted length of as much as approximately 3.6 tokens and a 2.6-fold decoding speedup in its test configuration.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

MiMo-V2.5 and V2.5-Pro retain three lightweight MTP layers.

### Multi-Teacher On-Policy Distillation

**Multi-Teacher On-Policy Distillation** (MOPD) is Xiaomi's principal capability-integration method for post-training the V2 family.<sup>[\[20\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MOPD-20)</sup>

The pipeline has three broad stages:

1.  supervised fine-tuning for general instruction following;
2.  independent reinforcement learning or specialist training for domain teachers;
3.  on-policy distillation from the specialist teachers into one student model.

Teacher domains can include:

- mathematics;
- code;
- agentic search;
- terminal work;
- general tool use;
- safety;
- general reasoning;
- instruction following.

Instead of training the student only on static teacher-generated answers, the student produces its own current rollouts. The relevant teacher then supplies token-level guidance for those student trajectories.

A simplified reverse-KL distillation loss is:

$$
\mathcal{L}_{distill} = {\mathbb{E}}_{y \sim \pi_{\theta}}\left\lbrack \sum\limits_{t}D_{KL}\left( \pi_{\theta}( \cdot \mid y_{< t},x) \parallel \pi_{{teacher}(x)}( \cdot \mid y_{< t},x) \right) \right\rbrack.
$$

Outcome rewards from executable verifiers or reward models can be combined with the token-level distillation signal.

MOPD is intended to reduce two problems:

- **capability interference** — improving one domain while degrading another;
- **distribution mismatch** — training on teacher trajectories that the student itself would rarely produce.

The teachers, student, reward functions, and environment distribution remain developer-selected, so MOPD does not guarantee that every integrated capability will equal the best specialist.

### Rollout Routing Replay

MoE reinforcement learning can become unstable when inference and training route the same token to different experts because of numerical precision, implementation, or synchronization differences.

Xiaomi introduced **Rollout Routing Replay** (R3), which records the expert-routing decisions used during rollout generation and reuses them when the trajectory is trained.<sup>[\[21\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-R3-21)</sup>

For multi-turn agent training, Xiaomi also used a request-level prefix cache containing:

- KV-cache state;
- previously selected experts;
- reusable request prefixes.

This improves rollout consistency but requires additional storage and coordination between the inference and training systems.

### Native multimodality in MiMo-V2.5

MiMo-V2.5 uses separate encoders connected to one language-and-agent backbone.

For an image or video input $\mathbf{X}$:

$$
\mathbf{Z}_{vis} = E_{vis}(\mathbf{X}),
$$

$$
\mathbf{H}_{vis} = P_{vis}(\mathbf{Z}_{vis}),
$$

where $E_{vis}$ is the visual encoder and $P_{vis}$ maps its output into the language-model representation space.

For audio $\mathbf{A}$:

$$
\mathbf{Z}_{aud} = E_{aud}(\mathbf{A}),\qquad\mathbf{H}_{aud} = P_{aud}(\mathbf{Z}_{aud}).
$$

The resulting visual, audio, and text representations are inserted into a shared sequence processed by the MoE language backbone.

The visual encoder has:

- approximately 729M parameters;
- 28 layers;
- 24 window-attention layers;
- four full-attention layers;
- 32 query heads and eight key–value heads;
- alternating row and column windows.

The audio encoder has:

- approximately 261M parameters;
- 24 layers;
- 12 sliding-window and 12 full-attention layers;
- a 128-token window;
- 16 query and 16 key–value heads.<sup>[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)</sup>

MiMo-V2.5 produces text or tool calls. It does not natively return generated images, video, or speech through the language-model endpoint. Speech output requires a TTS model.

### Long-context operation

MiMo-V2.5 and V2.5-Pro accept up to approximately one million tokens in their post-trained configurations. This enables workloads involving:

- large repositories;
- long agent histories;
- multiple documents;
- lengthy audio or video representations;
- extensive tool output;
- multi-stage project memory.

A nominal one-million-token window does not establish:

- perfect recall at every location;
- uniform accuracy as prompt length increases;
- correct resolution of contradictory instructions;
- immunity to prompt injection;
- reliable preservation of all state;
- constant cost or latency.

In Xiaomi's GraphWalks evaluation, MiMo-V2.5-Pro retained non-zero graph-reasoning performance at one million tokens, while the earlier V2-Pro checkpoint degraded sharply. The remaining gap from perfect accuracy illustrates the distinction between accepted context length and effective use of that context.<sup>[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

### UltraSpeed inference

The UltraSpeed configuration combines three major techniques.<sup>[\[18\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-UltraSpeed-18)</sup>

**Selective FP4 quantization.** MoE expert weights are converted to MXFP4, while more sensitive attention, embedding, routing, and other components remain at higher precision.

Approximate raw storage for $N$ parameters at $b$ bits is:

$$
{Storage} \approx N\frac{b}{8}.
$$

At four bits, 1.02 trillion weights correspond to approximately 510 GB before higher-precision modules, scales, buffers, and runtime memory.

**DFlash speculative decoding.** A masked block predictor proposes several tokens in parallel. The released configuration uses blocks of up to eight tokens. Xiaomi reported average accepted lengths of approximately:

- 6.30 for coding;
- 5.56 for mathematics and reasoning;
- 4.29 for agent tasks.

**TileRT kernels.** Persistent kernels retain the execution pipeline on the GPU and reduce repeated kernel-launch and synchronization overhead. Warp specialization separates communication, memory movement, and tensor operations into coordinated execution roles.

Xiaomi reported more than 1,000 output tokens per second and peaks near 1,200 tokens per second on one standard eight-GPU node. These are serving measurements for a specific quantization, speculative-decoding system, workload, and hardware configuration—not a universal speed of the original FP8 checkpoint.

## Pre-training and post-training

### MiMo-7B pre-training

MiMo-7B-Base was trained on approximately 25 trillion tokens in three broad stages.<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

The data mixture included:

- web text;
- books;
- academic and technical material;
- mathematics;
- source code;
- synthetic reasoning examples;
- creative-writing data.

The first stages used an 8,192-token sequence length. During the final stage, context was extended to 32,768 tokens and approximately 10% synthetic responses for mathematics, code, and creative-writing prompts were incorporated.

The preprocessing pipeline attempted to increase the density of useful reasoning patterns through improved extraction, multi-dimensional filtering, and synthetic-data generation.

The complete source inventory and document-level provenance were not released.

### MiMo-7B post-training

The initial reinforcement-learning collection contained approximately:

- 100,000 mathematics problems;
- 30,000 programming problems.

Rewards were primarily rule-based. Mathematics answers could be checked symbolically or against exact targets, while generated programs were compiled and executed against tests.

Xiaomi introduced a test-difficulty-driven code reward intended to avoid assigning identical credit to models that passed only simple tests and models that solved harder cases.

The report also discussed:

- reward hacking in mathematics;
- language mixing;
- interference between mathematics and coding;
- instability in vanilla GRPO;
- improved results from larger supervised-fine-tuning datasets;
- extended generation budgets during on-policy reinforcement learning.

These observations show that a higher reinforcement-learning reward is not automatically equivalent to better general reasoning.

### MiMo-VL training

MiMo-VL used approximately 2.4 trillion training tokens distributed across four stages:<sup>[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)</sup>

| Stage             | Purpose                          | Approximate tokens | Sequence length |                           |                                    |      |     |                                 |                                                         |      |     |                  |                                                                       |      |     |
|-------------------|----------------------------------|--------------------|-----------------|---------------------------|------------------------------------|------|-----|---------------------------------|---------------------------------------------------------|------|-----|------------------|-----------------------------------------------------------------------|------|-----|
| Projector warm-up | Initial image–language alignment | 300B               | 8K              | Vision–language alignment | Train visual encoder and projector | 167B | 8K  | General multimodal pre-training | Images, video, OCR, grounding, GUI, text, and reasoning | 1.4T | 8K  | Long-context SFT | Documents, high-resolution images, extended video, and long reasoning | 550B | 32K |

Post-training used MORL to combine:

- exact or rule-verifiable visual rewards;
- grounding and localization rewards;
- mathematical and logical rewards;
- learned human-preference rewards.

The paper reported that simultaneous optimization remained difficult because perception, grounding, and long-form reasoning could improve at different rates.

### MiMo-V2-Flash pre-training

MiMo-V2-Flash was trained on approximately 27 trillion tokens using FP8 mixed precision.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

| Stage                | Token range | Main objective                                                |                         |        |                                                              |                   |        |                                                                   |
|----------------------|-------------|---------------------------------------------------------------|-------------------------|--------|--------------------------------------------------------------|-------------------|--------|-------------------------------------------------------------------|
| Initial pre-training | 0–22T       | General text, code, mathematics, and knowledge at 32K context | Capability mid-training | 22–26T | Upsampled code and approximately 5% synthetic reasoning data | Context extension | 26–27T | Extension to 256K and greater emphasis on long-range dependencies |

The model uses 48 layers, of which 39 use sliding-window attention and nine use global attention. All layers except the first contain sparse MoE feed-forward components.

The training stack retained selected operations in BF16 or FP32 while using FP8 for many matrix multiplications. Router parameters were kept at higher precision.

### Agentic post-training

MiMo-V2-Flash and later models were trained in interactive environments including:

- code repositories;
- terminals;
- web development;
- browser search;
- general tool use;
- structured-function calls;
- multi-turn agents.

The V2-Flash report describes on-policy code-agent reinforcement learning across approximately 120,000 interactive environments.<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

The reported infrastructure combines:

- SGLang for rollout inference;
- Megatron-LM for training;
- FP8 training and rollout execution;
- R3 router replay;
- a data scheduler;
- a tool manager and toolbox;
- request-level prefix caching.

Executable environments provide stronger verification than subjective text-only evaluation, but they can contain faulty tests, incomplete requirements, or exploitable reward conditions.

### MiMo-V2.5 training

MiMo-V2.5 was trained on approximately 48 trillion tokens through five stages:<sup>[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)</sup>

1.  text pre-training of the language backbone;
2.  warm-up of the visual and audio projectors;
3.  large-scale multimodal pre-training;
4.  supervised fine-tuning and agentic post-training, with context extension from 32K to 256K and one million tokens;
5.  reinforcement learning and MOPD.

The complete 48-trillion-token figure covers several modalities and stages. It should not be interpreted as 48 trillion independently authored text tokens.

### MiMo-V2.5-Pro training

MiMo-V2.5-Pro was pretrained on approximately 27 trillion tokens in FP8 mixed precision, with a native 32K sequence length and later extension to one million tokens.<sup>[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

Its post-training used:

- supervised fine-tuning;
- domain-specialized reinforcement learning;
- agentic reinforcement learning;
- MOPD integration.

Xiaomi has not released a complete source-by-source training-corpus inventory, every reward function, or all production agent environments.

## Main model families

### Summary table

| Model                        | Release       | Total / active parameters                    | Maximum published context                            | Native inputs                         | Primary role                                  |
|------------------------------|---------------|----------------------------------------------|------------------------------------------------------|---------------------------------------|-----------------------------------------------|
| **MiMo-7B**                  | May 2025      | Approximately 7B dense                       | Approximately 32K; later reasoning checkpoint at 48K | Text                                  | Reasoning, mathematics, and code              |
| **MiMo-VL-7B**               | June 2025     | Approximately 7B plus visual encoder         | Approximately 32K                                    | Text, images, and sampled video       | Multimodal reasoning and GUI grounding        |
| **MiMo-Audio-7B**            | Late 2025     | 7B-class language model plus audio tokenizer | Checkpoint-specific                                  | Text and audio                        | Audio understanding and generation            |
| **MiMo-Embodied-7B**         | November 2025 | Approximately 7B                             | Checkpoint-specific                                  | Text and visual/embodied observations | Driving, robotics, and planning               |
| **MiMo-V2-Flash**            | December 2025 | 309B / 15B                                   | 256K                                                 | Text                                  | Efficient reasoning, coding, and agents       |
| **MiMo-V2-Pro**              | March 2026    | More than 1T / 42B                           | 1M                                                   | Text                                  | Hosted flagship agent model, later superseded |
| **MiMo-V2-Omni**             | March 2026    | Not disclosed                                | Approximately 256K hosted                            | Text, images, video, and audio        | Omni-modal perception and agents              |
| **MiMo-V2.5**                | April 2026    | 310B / 15B                                   | 1M post-trained; 256K base                           | Text, images, video, and audio        | Current omni-modal agent model                |
| **MiMo-V2.5-Pro**            | April 2026    | 1.02T / 42B                                  | 1M post-trained; 256K base                           | Text                                  | Current flagship language and agent model     |
| **MiMo-V2.5-Pro-UltraSpeed** | June 2026     | Same foundation; selective FP4               | Service-specific                                     | Text                                  | High-throughput serving configuration         |

<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)[\[12\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoAudio-12)[\[13\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoEmbodied-13)[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)[\[15\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Pro-15)[\[16\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Omni-16)[\[5\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Card-5)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)[\[18\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-UltraSpeed-18)</sup>

### MiMo-7B-Base

MiMo-7B-Base is the initial dense research checkpoint. Its importance lies less in product deployment than in Xiaomi's investigation of how pre-training data can preserve reasoning potential before reinforcement learning.

The report found that the base checkpoint could improve substantially under direct reinforcement learning, sometimes exceeding larger comparison base models during the training process.<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

### MiMo-7B-RL-Zero

MiMo-7B-RL-Zero was trained directly from the base model without a conventional instruction-tuned cold start.

This approach encouraged exploration but also produced:

- unstable output formatting;
- language mixing;
- different learning rates across mathematics and code;
- greater opportunity for reward exploitation.

The model was released primarily as a research checkpoint.

### MiMo-7B-SFT and MiMo-7B-RL

MiMo-7B-SFT added instruction and reasoning examples before reinforcement learning. The subsequent MiMo-7B-RL model achieved higher final scores than RL-Zero on the reported mathematics and coding evaluations.

Later experiments increased supervised data from approximately 500,000 examples to six million and extended the reinforcement-learning generation budget to as much as 48K tokens.

### MiMo-VL-7B

MiMo-VL-7B uses Qwen2.5-ViT as its original visual encoder, a learned MLP projector, and the MiMo-7B language backbone.<sup>[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)</sup>

Its intended tasks include:

- image question answering;
- mathematical diagrams;
- document and chart analysis;
- OCR;
- temporal localization in video;
- GUI element grounding;
- screen-based agent actions.

The model can output an explicit reasoning stream. That stream is generated text and should not be treated as a complete record of every internal computation.

### MiMo-Audio-7B

MiMo-Audio uses a learned audio tokenizer with approximately 1.2 billion parameters. The tokenizer operates at approximately 25 time steps per second and uses eight residual-vector-quantization codebooks. This corresponds to approximately 200 discrete audio tokens per second before patch compression.<sup>[\[12\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoAudio-12)</sup>

A patch encoder aggregates four time steps, reducing the sequence seen by the main language model to approximately 6.25 patches per second. A patch decoder expands language-model outputs back into the full multi-codebook representation.

The system supports tasks in which the input and output can contain different combinations of audio and text.

### MiMo-V2-Flash

MiMo-V2-Flash is the architectural foundation for the V2.5 language backbone. It was designed to minimize active parameters and KV-cache cost without reducing the total expert capacity as aggressively.

The model supports thinking and non-thinking response styles through chat templates and post-training. Its downloadable weights include Multi-Token Prediction components.

### MiMo-V2-Pro

MiMo-V2-Pro was the first hosted trillion-parameter MiMo agent model. It used more than one trillion total parameters, 42 billion active parameters, a seven-to-one local/global attention ratio, and a one-million-token context.<sup>[\[15\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Pro-15)</sup>

It was not released with the same complete open-weight package later provided for V2.5-Pro.

### MiMo-V2-Omni

MiMo-V2-Omni was a hosted omni-modal model with dedicated visual and audio encoders and a shared reasoning backbone.<sup>[\[16\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2Omni-16)</sup>

Xiaomi demonstrated:

- long-form audio understanding;
- audio–video joint reasoning;
- graphical-interface operation;
- browser automation;
- future-event prediction from video;
- structured tool calls.

Claims such as understanding more than ten hours of continuous audio were developer demonstrations, not proof of uniform reliability over all ten-hour recordings.

### MiMo-V2.5

MiMo-V2.5 consolidates the V2-Flash language architecture and V2-Omni perception branch into one downloadable model.

It supports:

- native text, image, video, and audio input;
- structured tool calls;
- long-context reasoning;
- GUI and browser tasks;
- audiovisual analysis;
- software engineering;
- agent trajectories.

The model's output remains text or structured actions. It does not replace a dedicated speech or media generator.

### MiMo-V2.5-Pro

MiMo-V2.5-Pro targets the highest-compute language and agent workloads.

Its release demonstrations included:

- a SysY-to-RISC-V compiler implemented in Rust over 672 tool calls and 4.3 hours;
- an 8,192-line desktop video editor produced over 1,868 tool calls and 11.5 hours;
- iterative analog-circuit design using an ngspice simulation loop.<sup>[\[2\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Pro-2)</sup>

These are selected case studies using specific harnesses. They do not establish a general success rate for arbitrary multi-hour tasks.

## Speech and audio models

### MiMo-V2-TTS

MiMo-V2-TTS uses a proprietary audio tokenizer and multi-codebook speech–text modelling architecture. Xiaomi reports pre-training on more than 100 million hours of speech data.<sup>[\[17\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2TTS-17)</sup>

The system supports:

- global speaking-style instructions;
- localized emotional changes;
- gradual emotional transitions;
- natural prosody;
- voice cloning;
- singing;
- reinforcement-learning optimization for audio quality and articulation.

The parameter count and complete production architecture were not disclosed.

### MiMo-V2.5-TTS

MiMo-V2.5-TTS provides curated stock voices and natural-language control over:

- pace;
- emotion;
- tone;
- delivery;
- character;
- scene context.

It also accepts screenplay-style inputs that separate character identity, scene description, stage direction, and spoken text.<sup>[\[6\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25TTS-6)</sup>

### MiMo-V2.5-TTS-VoiceDesign

VoiceDesign creates a new synthetic voice from a textual description without requiring reference audio.

Descriptions can specify characteristics such as:

- approximate age;
- accent;
- pitch;
- timbre;
- speaking speed;
- temperament;
- vocal texture;
- performance style.

The generated result is a synthetic interpretation of the description, not a stable biometric identity unless the surrounding service separately preserves one.

### MiMo-V2.5-TTS-VoiceClone

VoiceClone reproduces a target voice from a short reference clip, reportedly within seconds and without a separate fine-tuning operation.<sup>[\[6\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25TTS-6)</sup>

The cloned voice can then be controlled through:

- natural-language directions;
- free-form inline audio tags;
- emotion and state tags;
- screenplay-style scripts.

Voice similarity does not establish the speaker's consent or the authenticity of the generated statement.

### MiMo-V2.5-ASR

MiMo-V2.5-ASR is an end-to-end speech-recognition model distributed under Apache License 2.0.<sup>[\[7\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ASR-7)</sup>

Its intended coverage includes:

- Mandarin Chinese;
- English;
- Wu;
- Cantonese;
- Hokkien;
- Sichuanese;
- Chinese–English code switching;
- song lyrics;
- noisy and far-field speech;
- overlapping speakers;
- technical and knowledge-intensive material;
- native punctuation.

The model uses MiMo-Audio-Tokenizer as its audio front end. Its training included large-scale mid-training, supervised fine-tuning, and reinforcement learning.

The model does not provide speaker diarization simply because it can recognize overlapping speech. Assigning words to named speakers remains a separate problem unless explicitly implemented by the surrounding system.

## Embodied and specialized models

### MiMo-Embodied

MiMo-Embodied attempts to share representations across multiple forms of embodiment.

Its task categories include:

- visual and spatial understanding;
- object affordance;
- task planning;
- future-state prediction;
- autonomous-driving perception;
- driving planning;
- robot action reasoning.

A shared foundation model can transfer concepts among environments, but it can also encounter conflicts between action spaces, sensor assumptions, and safety constraints.

MiMo-Embodied's outputs require a control framework before they can operate a physical system. A generated action description is not itself a validated control command.

### Smart-home and edge models

Xiaomi researchers have also published MiMo-derived models for smart-home perception and edge deployment, including systems that recognize gestures, activities, and household situations.

These specialist models may use smaller checkpoints and task-specific post-training. They should not be assigned the parameter count or one-million-token context of the V2.5 flagship models.

### Research architectures

Xiaomi has published research on architectures related to future MiMo development.

**HySparse** interleaves full-attention layers with several sparse-attention layers. Sparse layers reuse both token selections and KV-cache information derived from a preceding full-attention layer.<sup>[\[22\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-HySparse-22)</sup>

This differs from MiMo-V2.5's fixed sliding-window attention. HySparse selects non-local tokens dynamically and is a research architecture rather than a published specification of the V2.5 checkpoint.

**ARL-Tangram** investigates resource-efficient agentic reinforcement learning, including scheduling and allocation of rollout and training resources. It is infrastructure research rather than a standalone user-facing model.

## Reported benchmark results

Benchmark results in this section are mainly developer-reported. They depend on:

- checkpoint version;
- prompt format;
- number of attempts;
- generation length;
- reasoning mode;
- tool access;
- context-management strategy;
- agent harness;
- judge model;
- benchmark revision;
- repository and dependency state.

Results from separate publications should not be treated as one controlled longitudinal experiment.

### MiMo-7B

Selected MiMo-7B-RL results from the May 2025 report are:

| Benchmark                                                                           | MiMo-7B-Base | MiMo-7B-RL-Zero | MiMo-7B-SFT | MiMo-7B-RL |
|-------------------------------------------------------------------------------------|--------------|-----------------|-------------|------------|
| [MATH-500](https://systems-analysis.info/eng/MATH-500 "MATH-500")                   | 37.4         | 93.6            | 93.0        | 95.8       |
| AIME 2024                                                                           | 32.9         | 56.4            | 58.7        | 68.2       |
| AIME 2025                                                                           | 24.3         | 46.3            | 44.3        | 55.4       |
| [LiveCodeBench](https://systems-analysis.info/eng/LiveCodeBench "LiveCodeBench") v5 | 32.9         | 49.1            | 52.3        | 57.8       |
| LiveCodeBench v6                                                                    | 29.1         | 42.9            | 45.5        | 49.3       |

<sup>[\[10\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7B-10)</sup>

A later checkpoint trained with six million supervised examples and a 48K reasoning budget reported 80.1 on AIME 2024 and 70.2 on AIME 2025. These results used longer generation and a later training recipe and are not directly equivalent to the original table.

### MiMo-VL-7B

Selected developer-reported results include:

| Benchmark       | MiMo-VL-7B-SFT | MiMo-VL-7B-RL           | Qualification                          |
|-----------------|----------------|-------------------------|----------------------------------------|
| MMMU validation | 64.6           | 66.7                    | General visual understanding           |
| OlympiadBench   | 59.4           | 59.4                    | Multimodal mathematical reasoning      |
| MathVision      | 57.9           | 60.4                    | Visual mathematics                     |
| MathVista       | —              | 81.8                    | Multimodal reasoning                   |
| CharXiv-RQ      | —              | 56.5                    | Scientific-chart understanding         |
| OSWorld-G       | —              | Approximately 54.7–56.1 | Version-dependent GUI-grounding report |

<sup>[\[11\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoVL-11)</sup>

The paper reports that MiMo-VL-7B-RL exceeded Qwen2.5-VL-7B on 35 of 40 selected tasks. The comparison was conducted under Xiaomi's evaluation configuration.

### MiMo-V2-Flash

Selected results reported for MiMo-V2-Flash include:

| Benchmark                                                                                         | MiMo-V2-Flash    |
|---------------------------------------------------------------------------------------------------|------------------|
| [SWE-bench Verified](https://systems-analysis.info/eng/SWE-bench_Verified "SWE-bench Verified")   | 73.4%            |
| SWE-bench Multilingual                                                                            | 71.7%            |
| AIME 2025                                                                                         | Approximately 94 |
| [GPQA-Diamond](https://systems-analysis.info/eng/GPQA_Diamond_Benchmark "GPQA Diamond Benchmark") | Approximately 84 |
| LiveCodeBench                                                                                     | Approximately 83 |
| Tau²-Bench                                                                                        | Approximately 80 |
| LongBench v2                                                                                      | 60.6             |

<sup>[\[14\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashReport-14)</sup>

Several of these tests used thinking mode, long outputs, tools, or an agent harness. SWE-bench results measure the combined model and software-engineering framework.

### MiMo-V2.5 and V2.5-Pro base models

The published base-model table reports:

| Benchmark                                                                                                      | MiMo-V2.5 Base | MiMo-V2.5-Pro Base |
|----------------------------------------------------------------------------------------------------------------|----------------|--------------------|
| [MMLU](https://systems-analysis.info/eng/MMLU_Benchmark "MMLU Benchmark"), 5-shot                              | 86.3           | 89.4               |
| MMLU-Redux, 5-shot                                                                                             | 89.8           | 92.8               |
| [MMLU-Pro](https://systems-analysis.info/eng/MMLU-Pro_Benchmark "MMLU-Pro Benchmark"), 5-shot                  | 65.8           | 68.5               |
| [GPQA-Diamond](https://systems-analysis.info/eng/GPQA_Diamond_Benchmark "GPQA Diamond Benchmark"), 5-shot      | 58.1           | 66.7               |
| [GSM8K](https://systems-analysis.info/eng/GSM8K_(Grade_School_Math_8K) "GSM8K (Grade School Math 8K)"), 8-shot | 83.3           | 99.6               |
| [MATH](https://systems-analysis.info/eng/MATH_Benchmark "MATH Benchmark"), 4-shot                              | 67.7           | 86.2               |
| HumanEval+                                                                                                     | 71.3           | 75.6               |
| LiveCodeBench v6                                                                                               | 35.5           | 39.6               |
| C-Eval, 5-shot                                                                                                 | 88.6           | 91.5               |
| CMMLU, 5-shot                                                                                                  | 88.2           | 90.2               |
| GlobalMMLU, 5-shot                                                                                             | 77.4           | 83.6               |

<sup>[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

These are base-checkpoint results without the complete agentic post-training applied to the main hosted models.

### Current agent evaluations

Xiaomi reports that MiMo-V2.5 scored 62.3 on the general subset of Claw-Eval and approached the token-efficiency frontier of the comparison systems.<sup>[\[4\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25-4)</sup>

For MiMo-V2.5-Pro, Xiaomi's release emphasized:

- higher internal MiMo Coding Bench scores;
- improved GraphWalks long-context performance;
- 64% Pass³ on Claw-Eval at approximately 70,000 tokens per trajectory;
- lower token use than selected competing frontier systems at comparable developer-reported capability.<sup>[\[2\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25Pro-2)[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

Internal MiMo Coding Bench is not a fully independent public evaluation. Comparative results involving unreleased or provider-controlled models may also change after server-side updates.

### Long-context evaluation

MiMo-V2.5-Pro was evaluated on GraphWalks from 32K to one million input tokens. At one million tokens, the reported scores were approximately:

- 0.37 on breadth-first-search queries;
- 0.62 on parent-listing queries.<sup>[\[3\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProCard-3)</sup>

These values are evidence of partial long-context reasoning rather than perfect retrieval. The earlier MiMo-V2-Pro service reportedly fell to zero at the longest tested setting.

## Products and ecosystem

### MiMo API

The Xiaomi MiMo API provides model access through interfaces compatible with:

- OpenAI-style chat and response formats;
- Anthropic-style Messages formats;
- function and tool calls;
- streaming generation;
- multimodal inputs;
- context caching;
- reasoning modes.

Protocol compatibility does not imply identical tokenization, supported parameters, system prompts, safety behaviour, or streaming events.

The V2.5 model identifiers include:

- `mimo-v2.5-pro`;
- `mimo-v2.5-pro-ultraspeed`, subject to access controls;
- `mimo-v2.5`;
- MiMo-V2.5 TTS identifiers;
- MiMo-V2.5-ASR.<sup>[\[8\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-APIModels-8)</sup>

Prices and access conditions are service settings that can change and are not model-architecture properties.

### MiMo Studio

MiMo Studio is Xiaomi's browser-based model interface. It supports model switching, multimodal input, and long-context interaction.

Product-level results can involve:

- file parsing;
- prompt templates;
- search;
- moderation;
- media processing;
- persistent conversation state.

These components should not automatically be attributed to the raw checkpoint.

### MiMo Code

MiMo Code is a software-engineering agent built around three themes:<sup>[\[23\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMoCode-23)</sup>

- computation;
- memory;
- evolution.

Its workflows can include:

- repository exploration;
- source-code editing;
- compilation;
- tests;
- shell commands;
- plans and checklists;
- long-running task memory;
- reusable skills;
- review and repair cycles.

MiMo Code is a system rather than one foundation checkpoint. Its reliability depends on the selected model, sandbox, permissions, memory implementation, and developer tooling.

### MiMo Claw

MiMo Claw is an agent platform integrating MiMo models with OpenClaw-style tools and productivity services.

It supports:

- multi-step task planning;
- file and document workflows;
- tool calls;
- web and application interaction;
- structured outputs;
- autonomous execution.

An agent platform must impose authorization boundaries independently of the model. A model deciding that an action is useful does not constitute user approval for a purchase, message, deletion, or disclosure.

### Xiaomi-device integration

MiMo research is relevant to Xiaomi's broader ecosystem of:

- smartphones;
- personal computers;
- smart-home devices;
- vehicles;
- wearables;
- robotics.

The existence of a model capability does not establish that it is deployed in every Xiaomi product. Device availability depends on product announcements, region, hardware, privacy controls, and software version.

### Third-party agent frameworks

Xiaomi publishes examples for using MiMo models with coding and agent scaffolds such as:

- Claude Code-compatible interfaces;
- OpenCode;
- Kilo;
- Cline;
- OpenClaw;
- SGLang and vLLM tool-call parsers.

The scaffold contributes prompts, tools, context management, and execution logic. A benchmark produced in one scaffold may not transfer unchanged to another.

## Deployment and resource requirements

### Supported inference software

Depending on the checkpoint, Xiaomi documents or provides examples for:

- SGLang;
- vLLM;
- Hugging Face Transformers;
- KTransformers;
- Unsloth;
- Docker Model Runner;
- community llama.cpp, Ollama, or LM Studio conversions.

New architectures may require recent development versions, custom model code, or specialized kernels.

### Model storage

Raw storage can be approximated by:

$$
{Bytes} \approx N_{parameters}\frac{bits}{8}.
$$

Ignoring metadata and higher-precision components:

| Model                | 16-bit weights        | 8-bit weights         | 4-bit weights        |
|----------------------|-----------------------|-----------------------|----------------------|
| MiMo-V2.5, 310B      | Approximately 620 GB  | Approximately 310 GB  | Approximately 155 GB |
| MiMo-V2.5-Pro, 1.02T | Approximately 2.04 TB | Approximately 1.02 TB | Approximately 510 GB |

Actual memory use is higher because of:

- quantization scales;
- higher-precision layers;
- multimodal encoders;
- routing buffers;
- KV cache;
- MTP state;
- CUDA graphs;
- workspaces;
- batch scheduling;
- framework overhead.

A 42-billion-active MoE model does not have the storage footprint of a dense 42-billion-parameter model.

### Expert parallelism

Large MiMo checkpoints require some combination of:

- tensor parallelism;
- expert parallelism;
- pipeline parallelism;
- data-parallel attention;
- high-bandwidth all-to-all communication;
- distributed loading.

Expert activation can be computationally sparse while communication remains expensive. Small batches may fail to use the distributed expert hardware efficiently.

### Long-context memory

Hybrid sliding-window attention reduces KV-cache growth but does not eliminate it. Global-attention layers still maintain long-range cache state.

One-million-token serving also increases:

- prompt-transfer time;
- tokenization time;
- prefill work;
- cache management;
- opportunity for irrelevant information;
- cost of replaying an agent history.

In practical systems, retrieval, summaries, structured memory, and selective context loading may be more effective than placing every available token into one prompt.

## Licensing

MiMo releases do not share one family-wide licence.

| Model or component               | Principal licence or status                                                             |
|----------------------------------|-----------------------------------------------------------------------------------------|
| MiMo-7B series                   | MIT License                                                                             |
| MiMo-VL-7B                       | MIT License                                                                             |
| MiMo-Audio models and tokenizer  | MIT License                                                                             |
| MiMo-V2-Flash                    | MIT License                                                                             |
| MiMo-V2.5 and MiMo-V2.5-Pro      | MIT License                                                                             |
| MiMo-V2.5-Pro-FP4-DFlash         | Checkpoint-specific open-weight terms; distributed through Xiaomi's official repository |
| MiMo-V2.5-ASR                    | Apache License 2.0                                                                      |
| MiMo-V2.5-TTS production models  | Proprietary hosted-service terms                                                        |
| MiMo Studio, Code, Claw, and API | Product and service terms                                                               |

<sup>[\[24\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-MiMo7BLicense-24)[\[25\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V2FlashLicense-25)[\[26\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25License-26)[\[27\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ProLicense-27)[\[7\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-V25ASR-7)</sup>

An open model licence governs the distributed files. It does not automatically grant rights to:

- private prompts or uploaded data;
- a third party's voice or likeness;
- copyrighted source material;
- trademarks;
- every generated output;
- proprietary product services.

The training datasets and production infrastructure are not fully published, so *open-weight* remains a useful distinction.

## Safety, privacy, and regulation

### Model-level and system-level safety

A deployed MiMo application can include safeguards at several layers:

- training-data filtering;
- supervised safety data;
- safety-specialized reinforcement-learning teachers;
- MOPD;
- prompt and output classifiers;
- tool permissions;
- sandboxing;
- account-level controls;
- abuse monitoring;
- human review.

Downloadable weights can be served without Xiaomi's product-level controls. Deployers must therefore perform independent evaluation and apply safeguards appropriate to the application.

### Agent security

Agent deployments should use:

- least-privilege credentials;
- isolated filesystems;
- network allowlists;
- command restrictions;
- secret separation;
- execution limits;
- transaction limits;
- audit logs;
- approval for irreversible actions;
- rollback procedures.

Model alignment is not a substitute for operating-system or application security.

### Prompt injection

Images, websites, source files, emails, documents, and tool outputs can contain instructions intended to override the user's goal.

Applications should distinguish:

- system instructions;
- developer instructions;
- user requests;
- trusted application state;
- retrieved evidence;
- untrusted content;
- executable commands.

A one-million-token context can increase the amount of untrusted content available to an attacker.

### Privacy

MiMo applications can process:

- source code;
- private documents;
- photographs;
- vehicle or home-camera video;
- voice samples;
- meeting recordings;
- location information;
- health or financial data.

Users and deployers should determine:

- where data are stored;
- whether data are logged;
- retention periods;
- whether content is used for model improvement;
- access permissions;
- cross-border transfers;
- deletion procedures.

Audio and facial data can constitute biometric or otherwise sensitive personal information.

### Chinese generative-AI regulation

Public-facing generative-AI services in China operate under rules including the **Interim Measures for the Management of Generative Artificial Intelligence Services**, effective from 15 August 2023. These measures address lawful data, personal information, security, and content obligations.<sup>[\[28\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-ChinaGenAI-28)</sup>

The **Measures for Labeling Artificial Intelligence-Generated Synthetic Content** took effect on 1 September 2025 and require applicable visible and metadata-based labels for generated text, images, audio, video, and virtual scenes.<sup>[\[29\]](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_note-ChinaLabels-29)</sup>

Deployments in other jurisdictions may also be subject to privacy, consumer-protection, copyright, biometric, medical, vehicle-safety, and artificial-intelligence regulation.

## External links

- <a href="https://mimo.xiaomi.com/" class="external text" rel="nofollow">Xiaomi MiMo official portal</a>
- <a href="https://mimo.mi.com/docs/en-US/welcome" class="external text" rel="nofollow">Xiaomi MiMo API platform</a>
- <a href="https://huggingface.co/XiaomiMiMo" class="external text" rel="nofollow">Xiaomi MiMo on Hugging Face</a>
- <a href="https://github.com/XiaomiMiMo" class="external text" rel="nofollow">Xiaomi MiMo on GitHub</a>
- <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro" class="external text" rel="nofollow">MiMo-V2.5-Pro model card</a>
- <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5" class="external text" rel="nofollow">MiMo-V2.5 model card</a>
- <a href="https://github.com/XiaomiMiMo/MiMo-V2-Flash" class="external text" rel="nofollow">MiMo-V2-Flash repository</a>
- <a href="https://github.com/XiaomiMiMo/MiMo-V2.5-ASR" class="external text" rel="nofollow">MiMo-V2.5-ASR repository</a>
- <a href="https://arxiv.org/abs/2505.07608" class="external text" rel="nofollow">MiMo-7B technical report</a>
- <a href="https://arxiv.org/abs/2506.03569" class="external text" rel="nofollow">MiMo-VL technical report</a>
- <a href="https://arxiv.org/abs/2601.02780" class="external text" rel="nofollow">MiMo-V2-Flash technical report</a>
- <a href="https://arxiv.org/abs/2606.30406" class="external text" rel="nofollow">MOPD paper</a>

## See also

- [Large language model](https://systems-analysis.info/eng/Large_language_model "Large language model")
- [Multimodal large language model](https://systems-analysis.info/eng/Multimodal_large_language_models "Multimodal large language models")
- [Mixture of experts](https://systems-analysis.info/eng/Mixture-of-Experts_(MoE) "Mixture-of-Experts (MoE)")

## Literature

- Xiaomi LLM-Core Team. *MiMo: Unlocking the Reasoning Potential of Language Model — From Pretraining to Posttraining*. arXiv:2505.07608, 2025.
- Xiaomi LLM-Core Team. *MiMo-VL Technical Report*. arXiv:2506.03569, 2025.
- Xiaomi LLM-Core Team. *MiMo-Audio: Audio Language Models are Few-Shot Learners*. arXiv:2512.23808, 2025.
- Xiaomi MiMo Team. *MiMo-Embodied: X-Embodied Foundation Model Technical Report*. arXiv:2511.16518, 2025.
- Xiaomi LLM-Core Team. *MiMo-V2-Flash Technical Report*. arXiv:2601.02780, 2026.
- Ma, W. et al. *Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers*. arXiv:2510.11370, 2025.
- Gao, Y. et al. *HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing*. arXiv:2602.03560, 2026.
- Ma, W. et al. *MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training*. arXiv:2606.30406, 2026.
- Vaswani, A. et al. *Attention Is All You Need*. Advances in Neural Information Processing Systems 30, 2017.
- Shazeer, N. et al. *Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer*. arXiv:1701.06538, 2017.
- Ainslie, J. et al. *GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints*. EMNLP, 2023.
- Leviathan, Y.; Kalman, M.; Matias, Y. *Fast Inference from Transformers via Speculative Decoding*. ICML, 2023.

## References

1.  <span id="cite_note-MiMoPortal-1">↑ <sup>[1.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoPortal_1-0)</sup> <sup>[1.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoPortal_1-1)</sup> Xiaomi MiMo. *Official model and research portal*. Accessed 10 July 2026. <a href="https://mimo.xiaomi.com/" class="external free" rel="nofollow">https://mimo.xiaomi.com/</a></span>
2.  <span id="cite_note-V25Pro-2">↑ <sup>[2.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Pro_2-0)</sup> <sup>[2.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Pro_2-1)</sup> <sup>[2.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Pro_2-2)</sup> <sup>[2.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Pro_2-3)</sup> <sup>[2.4](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Pro_2-4)</sup> Xiaomi MiMo Team (2026). *Xiaomi MiMo-V2.5-Pro*, 27 April 2026. <a href="https://mimo.xiaomi.com/mimo-v2-5-pro/" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-5-pro/</a></span>
3.  <span id="cite_note-V25ProCard-3">↑ <sup>[3.00](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-0)</sup> <sup>[3.01](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-1)</sup> <sup>[3.02](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-2)</sup> <sup>[3.03](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-3)</sup> <sup>[3.04](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-4)</sup> <sup>[3.05](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-5)</sup> <sup>[3.06](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-6)</sup> <sup>[3.07](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-7)</sup> <sup>[3.08](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-8)</sup> <sup>[3.09](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-9)</sup> <sup>[3.10](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProCard_3-10)</sup> Xiaomi MiMo. *XiaomiMiMo/MiMo-V2.5-Pro*. Hugging Face model card. <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro</a></span>
4.  <span id="cite_note-V25-4">↑ <sup>[4.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25_4-0)</sup> <sup>[4.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25_4-1)</sup> <sup>[4.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25_4-2)</sup> <sup>[4.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25_4-3)</sup> Xiaomi MiMo Team (2026). *Xiaomi MiMo-V2.5*, 22 April 2026. <a href="https://mimo.xiaomi.com/mimo-v2-5/" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-5/</a></span>
5.  <span id="cite_note-V25Card-5">↑ <sup>[5.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-0)</sup> <sup>[5.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-1)</sup> <sup>[5.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-2)</sup> <sup>[5.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-3)</sup> <sup>[5.4](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-4)</sup> <sup>[5.5](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-5)</sup> <sup>[5.6](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25Card_5-6)</sup> Xiaomi MiMo. *XiaomiMiMo/MiMo-V2.5*. Hugging Face model card. <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-V2.5</a></span>
6.  <span id="cite_note-V25TTS-6">↑ <sup>[6.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25TTS_6-0)</sup> <sup>[6.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25TTS_6-1)</sup> <sup>[6.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25TTS_6-2)</sup> Xiaomi MiMo Team (2026). *MiMo-V2.5-TTS Series*. <a href="https://mimo.xiaomi.com/mimo-v2-5-tts" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-5-tts</a></span>
7.  <span id="cite_note-V25ASR-7">↑ <sup>[7.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ASR_7-0)</sup> <sup>[7.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ASR_7-1)</sup> <sup>[7.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ASR_7-2)</sup> Xiaomi MiMo Team (2026). *MiMo-V2.5-ASR: Robust Speech Recognition Across Languages, Dialects, and Complex Acoustic Scenarios*. <a href="https://github.com/XiaomiMiMo/MiMo-V2.5-ASR" class="external free" rel="nofollow">https://github.com/XiaomiMiMo/MiMo-V2.5-ASR</a></span>
8.  <span id="cite_note-APIModels-8">↑ <sup>[8.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-APIModels_8-0)</sup> <sup>[8.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-APIModels_8-1)</sup> Xiaomi MiMo. *API model catalogue*. Accessed 10 July 2026. <a href="https://mimo.mi.com/docs/en-US/quick-start/model" class="external free" rel="nofollow">https://mimo.mi.com/docs/en-US/quick-start/model</a></span>
9.  <span id="cite_note-ModelUpdates-9">↑ <sup>[9.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-ModelUpdates_9-0)</sup> <sup>[9.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-ModelUpdates_9-1)</sup> <sup>[9.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-ModelUpdates_9-2)</sup> Xiaomi MiMo. *Model release notes*. Accessed 10 July 2026. <a href="https://mimo.mi.com/docs/en-US/updates/model" class="external free" rel="nofollow">https://mimo.mi.com/docs/en-US/updates/model</a></span>
10. <span id="cite_note-MiMo7B-10">↑ <sup>[10.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-0)</sup> <sup>[10.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-1)</sup> <sup>[10.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-2)</sup> <sup>[10.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-3)</sup> <sup>[10.4](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-4)</sup> <sup>[10.5](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-5)</sup> <sup>[10.6](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-6)</sup> <sup>[10.7](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7B_10-7)</sup> Xiaomi LLM-Core Team (2025). *MiMo: Unlocking the Reasoning Potential of Language Model — From Pretraining to Posttraining*. arXiv:2505.07608. <a href="https://arxiv.org/abs/2505.07608" class="external free" rel="nofollow">https://arxiv.org/abs/2505.07608</a></span>
11. <span id="cite_note-MiMoVL-11">↑ <sup>[11.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-0)</sup> <sup>[11.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-1)</sup> <sup>[11.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-2)</sup> <sup>[11.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-3)</sup> <sup>[11.4](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-4)</sup> <sup>[11.5](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoVL_11-5)</sup> Xiaomi LLM-Core Team (2025). *MiMo-VL Technical Report*. arXiv:2506.03569. <a href="https://arxiv.org/abs/2506.03569" class="external free" rel="nofollow">https://arxiv.org/abs/2506.03569</a></span>
12. <span id="cite_note-MiMoAudio-12">↑ <sup>[12.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoAudio_12-0)</sup> <sup>[12.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoAudio_12-1)</sup> <sup>[12.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoAudio_12-2)</sup> <sup>[12.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoAudio_12-3)</sup> Xiaomi LLM-Core Team (2025). *MiMo-Audio: Audio Language Models are Few-Shot Learners*. arXiv:2512.23808. <a href="https://arxiv.org/abs/2512.23808" class="external free" rel="nofollow">https://arxiv.org/abs/2512.23808</a></span>
13. <span id="cite_note-MiMoEmbodied-13">↑ <sup>[13.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoEmbodied_13-0)</sup> <sup>[13.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoEmbodied_13-1)</sup> <sup>[13.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoEmbodied_13-2)</sup> Xiaomi MiMo Team (2025). *MiMo-Embodied: X-Embodied Foundation Model Technical Report*. arXiv:2511.16518. <a href="https://arxiv.org/abs/2511.16518" class="external free" rel="nofollow">https://arxiv.org/abs/2511.16518</a></span>
14. <span id="cite_note-V2FlashReport-14">↑ <sup>[14.00](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-0)</sup> <sup>[14.01](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-1)</sup> <sup>[14.02](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-2)</sup> <sup>[14.03](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-3)</sup> <sup>[14.04](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-4)</sup> <sup>[14.05](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-5)</sup> <sup>[14.06](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-6)</sup> <sup>[14.07](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-7)</sup> <sup>[14.08](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-8)</sup> <sup>[14.09](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-9)</sup> <sup>[14.10](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashReport_14-10)</sup> Xiaomi LLM-Core Team (2026). *MiMo-V2-Flash Technical Report*. arXiv:2601.02780. <a href="https://arxiv.org/abs/2601.02780" class="external free" rel="nofollow">https://arxiv.org/abs/2601.02780</a></span>
15. <span id="cite_note-V2Pro-15">↑ <sup>[15.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Pro_15-0)</sup> <sup>[15.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Pro_15-1)</sup> <sup>[15.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Pro_15-2)</sup> <sup>[15.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Pro_15-3)</sup> Xiaomi MiMo Team (2026). *Xiaomi MiMo-V2-Pro*, 18 March 2026. <a href="https://mimo.xiaomi.com/mimo-v2-pro" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-pro</a></span>
16. <span id="cite_note-V2Omni-16">↑ <sup>[16.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Omni_16-0)</sup> <sup>[16.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Omni_16-1)</sup> <sup>[16.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Omni_16-2)</sup> <sup>[16.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2Omni_16-3)</sup> Xiaomi MiMo Team (2026). *Xiaomi MiMo-V2-Omni*, 18 March 2026. <a href="https://mimo.xiaomi.com/mimo-v2-omni" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-omni</a></span>
17. <span id="cite_note-V2TTS-17">↑ <sup>[17.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2TTS_17-0)</sup> <sup>[17.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2TTS_17-1)</sup> <sup>[17.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2TTS_17-2)</sup> Xiaomi MiMo Team (2026). *Xiaomi MiMo-V2-TTS*, 18 March 2026. <a href="https://mimo.xiaomi.com/mimo-v2-tts" class="external free" rel="nofollow">https://mimo.xiaomi.com/mimo-v2-tts</a></span>
18. <span id="cite_note-UltraSpeed-18">↑ <sup>[18.0](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-UltraSpeed_18-0)</sup> <sup>[18.1](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-UltraSpeed_18-1)</sup> <sup>[18.2](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-UltraSpeed_18-2)</sup> <sup>[18.3](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-UltraSpeed_18-3)</sup> Xiaomi MiMo Team (2026). *MiMo-V2.5-Pro-UltraSpeed: Pushing 1T-Parameter Model Generation Speed to 1000 TPS*, 8 June 2026. <a href="https://mimo.xiaomi.com/blog/mimo-tilert-1000tps" class="external free" rel="nofollow">https://mimo.xiaomi.com/blog/mimo-tilert-1000tps</a></span>
19. <span id="cite_note-GQA-19">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-GQA_19-0) Ainslie, J. et al. (2023). *GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints*. Proceedings of EMNLP 2023. <a href="https://arxiv.org/abs/2305.13245" class="external free" rel="nofollow">https://arxiv.org/abs/2305.13245</a></span>
20. <span id="cite_note-MOPD-20">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MOPD_20-0) Ma, W. et al. (2026). *MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training*. arXiv:2606.30406. <a href="https://arxiv.org/abs/2606.30406" class="external free" rel="nofollow">https://arxiv.org/abs/2606.30406</a></span>
21. <span id="cite_note-R3-21">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-R3_21-0) Ma, W. et al. (2025). *Stabilizing MoE Reinforcement Learning by Aligning Training and Inference Routers*. arXiv:2510.11370. <a href="https://arxiv.org/abs/2510.11370" class="external free" rel="nofollow">https://arxiv.org/abs/2510.11370</a></span>
22. <span id="cite_note-HySparse-22">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-HySparse_22-0) Gao, Y. et al. (2026). *HySparse: A Hybrid Sparse Attention Architecture with Oracle Token Selection and KV Cache Sharing*. arXiv:2602.03560. <a href="https://arxiv.org/abs/2602.03560" class="external free" rel="nofollow">https://arxiv.org/abs/2602.03560</a></span>
23. <span id="cite_note-MiMoCode-23">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMoCode_23-0) Xiaomi MiMo Team (2026). *MiMo Code: Scaling Coding Agents to Long-Horizon Tasks*, 10 June 2026. <a href="https://mimo.xiaomi.com/blog/mimo-code-long-horizon" class="external free" rel="nofollow">https://mimo.xiaomi.com/blog/mimo-code-long-horizon</a></span>
24. <span id="cite_note-MiMo7BLicense-24">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-MiMo7BLicense_24-0) Xiaomi MiMo. *MiMo-7B License*. <a href="https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/blob/main/LICENSE" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-7B-Base/blob/main/LICENSE</a></span>
25. <span id="cite_note-V2FlashLicense-25">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V2FlashLicense_25-0) Xiaomi MiMo. *MiMo-V2-Flash License*. <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/blob/main/LICENSE" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/blob/main/LICENSE</a></span>
26. <span id="cite_note-V25License-26">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25License_26-0) Xiaomi MiMo. *MiMo-V2.5 License*. <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5/blob/main/LICENSE" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-V2.5/blob/main/LICENSE</a></span>
27. <span id="cite_note-V25ProLicense-27">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-V25ProLicense_27-0) Xiaomi MiMo. *MiMo-V2.5-Pro License*. <a href="https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/blob/main/LICENSE" class="external free" rel="nofollow">https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro/blob/main/LICENSE</a></span>
28. <span id="cite_note-ChinaGenAI-28">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-ChinaGenAI_28-0) Cyberspace Administration of China et al. (2023). *Interim Measures for the Management of Generative Artificial Intelligence Services*. <a href="https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm" class="external free" rel="nofollow">https://www.cac.gov.cn/2023-07/13/c_1690898327029107.htm</a></span>
29. <span id="cite_note-ChinaLabels-29">[↑](https://systems-analysis.info/eng/MiMo_(Xiaomi)#cite_ref-ChinaLabels_29-0) Cyberspace Administration of China et al. (2025). *Measures for Labeling Artificial Intelligence-Generated Synthetic Content*. <a href="https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm" class="external free" rel="nofollow">https://www.cac.gov.cn/2025-03/14/c_1743654684782215.htm</a></span>
