\n\n \n\n### 🔥🔥🔥 MiMo-VL 2508 Updates\n\nWe're excited to announce improvements to our MiMo-VL ([MiMo-VL-7B-RL-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508) and [MiMo-VL-7B-SFT-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-SFT-2508)), featuring enhanced performance across multiple benchmarks, improved thinking control capabilities, and better user experience.\n\n#### 📈 Performance Improvements\n\nMiMo-VL-7B-RL-2508 demonstrates consistent improvements across both image and video benchmarks, achieving notable milestones of **70.6 on MMMU** and **70.8 on VideoMME**.\n\n\n\nFull evaluation results can be found [below](#full-evaluation-results).\n\n#### 🤔 Thinking Control Feature\n\nA thinking control capability that allows users to turn off the model's reasoning mode using the no_think parameter:\n- Thinki"},{"ref":"P2","kind":"page","title":"XiaomiMiMo/MiMo repository metadata","date":"2026-06-11T03:20:37.087985+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo\n\nDescription: MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 2167\n\nForks: 94\n\nOpen issues: 38\n\nCreated: 2025-04-26T09:31:17Z\n\nPushed: 2025-06-05T16:01:49Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n---\n\n## Updates\n\n[2025.05.30] We scaled the SFT dataset from approximately 500K to 6M instances and continuously expanding the RL training window size from 32K to 48K, the performance of [MiMo-7B-RL-0530](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530) on AIME24 can be continuously improved and eventually surpass that of DeepSeek R1 (79.8).\n\n
\n\n \n\n## Introduction\n\nWelcome to the **MiMo-Audio-Eval** toolkit! This toolkit is designed to evaluate various audio language models as described in the **MiMo-Audio** paper. It provides a flexible and extensible framework, supporting a wide range of datasets, tasks, and models, specifically for evaluating pre-trained or supervised fine-tuned (SFT) audio language models. The toolkit is ideal for researchers and developers who need to assess the performance of these models across different tasks and datasets.\n\n## Supported Datasets, Tasks, and Models\n\nThe MiMo-Audio-Eval toolkit supports a comprehensive set of datasets, tasks, and models. Some of the key features include:\n\n* **Datasets**:\n\n* AISHELL1\n* LibriSpeech\n* SeedTTS\n* Expresso\n* InstructTTSEval\n* SpeechMMLU\n* MMAR\n* MMAU\n* MMAU-Pro\n* MMSU\n* ESD\n* Big Bench Audio\n* MultiChallenge "},{"ref":"P4","kind":"page","title":"XiaomiMiMo/MiMo-Audio-Tokenizer repository metadata","date":"2026-06-11T03:20:36.926389+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Audio-Tokenizer","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Audio-Tokenizer\n\nDescription: A unified tokenizer that is capable of both extracting semantic information and enabling high-fidelity audio reconstruction.\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 144\n\nForks: 13\n\nOpen issues: 2\n\nCreated: 2025-09-18T16:02:05Z\n\nPushed: 2025-09-19T08:11:24Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n\n\n# MiMo-Audio-Tokenizer\n\n\n\n
A unified tokenizer that is capable of both extracting semantic information and enabling high-fidelity audio reconstruction.
\n\n
\n\n## Key Features\n\n- Scaled parameters and training data bootstrap the frontier of audio tokenization\n- 1.2B pure transformer-based architecture to keep both efficiency and effectiveness\n- trained from scratch over 11 million hours covering both audio reconstruction task and the audio-to-text (A2T) task\n\n- Unified representation enhance both cross-modal alignment and speech reconstruction quality\n- jointly capture both semantic and acoustic information while further alleviates the semantic-acoustic representation conflict\n\n## Installation\n\n```sh\ngit clone https://github.com/XiaomiMiMo/MiMo-Audio-Tokenizer\ncd MiMo-Audio-Tokenizer\n# Install base dependencies\npip install -e .\n# Install flash-attn\npip install -e \".[flash]\"\n```\n\n## Model Download\n\n```sh\n# you might need `sudo apt-get install git-lfs` before download this model\ngit clone https://huggingface.co/XiaomiMiMo/MiMo-Audio-Tokenizer\n```\n\n## Example Usage\n\n### 0. Quick start\n\n```py\nimport torchaudio\nimport mimo_audio_tokenizer\n\n# one-line model init\ntokenizer = mimo_audio_tokenizer.load_model(\"path to your model\").bfloat16().cuda() # FlashAttention only support fp16 and bf16 data type\n\n# preprocess\nmels = []\nwav_paths = [\"mimo_audio_tokenizer/assets/BAC009S0764W0121.wav\", \"mimo_audio_tokenizer/assets/BAC009S0764W0122.wav\", \"mimo_audio_tokenizer/assets/猪八戒_gt.wav\"]\nfor wav_path in wav_paths:\nwav = mimo_audio_tokenizer.load_audio(wav_path, tokenizer"},{"ref":"P5","kind":"page","title":"XiaomiMiMo/MiMo-Audio repository metadata","date":"2026-06-11T03:20:36.753432+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Audio","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Audio\n\nDescription: MiMo-Audio: Audio Language Models are Few-Shot Learners\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 1046\n\nForks: 102\n\nOpen issues: 39\n\nCreated: 2025-09-19T00:46:49Z\n\nPushed: 2026-03-03T02:34:35Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nMiMo Audio: Audio Language Models are Few-Shot Learners\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n## Introduction\n\nExisting audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both"},{"ref":"P6","kind":"page","title":"XiaomiMiMo/MiMo-Audio-Training repository metadata","date":"2026-06-11T03:20:36.405358+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Audio-Training","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Audio-Training\n\nLanguage: Python\n\nStars: 109\n\nForks: 13\n\nOpen issues: 5\n\nCreated: 2025-10-16T13:52:54Z\n\nPushed: 2025-10-16T13:55:07Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n \n\n## Introduction\n\nWelcome to the **MiMo-Audio-Training** toolkit! This toolkit is designed to fine-tune the [XiaomiMiMo/MiMo-Audio-7B-Instruct](https://huggingface.co/XiaomiMiMo/MiMo-Audio-7B-Instruct). This toolkit serves as a reference implementation for researchers and developers interested in MiMo-Audio and looking to adapt it to their own custom tasks.\n\n## Supported Tasks\n\nThe MiMo-Audio-Eval toolkit supports a comprehensive set of tasks. Some of the key features include:\n\n* **Tasks**:\n\n* **SFT**:\n\n* ASR\n* TTS / InstructTTS\n* Audio Understanding and Reasoning\n* Spoken Dialogue\n\n## Getting Started\n\nTo get started with the MiMo-Audio-Training toolkit, follow the instructions below to set up the environment and install the required dependencies.\n\n### Prerequisites (Linux)\n\n* Python 3.12\n* CUDA >= 12.0\n\n### Installation:\n\n```bash\ngit clone --"},{"ref":"P7","kind":"page","title":"XiaomiMiMo/MiMo-Audio-Demo repository metadata","date":"2026-06-11T03:20:36.254511+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Audio-Demo","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Audio-Demo\n\nLanguage: HTML\n\nStars: 11\n\nForks: 3\n\nOpen issues: 1\n\nCreated: 2025-09-19T01:01:53Z\n\nPushed: 2025-11-17T13:34:24Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n# MiMo-Audio Blog\n\n- github: https://github.com/XiaomiMiMo/MiMo-Audio\n- gitpage: https://xiaomimimo.github.io/MiMo-Audio-Demo/"},{"ref":"P8","kind":"page","title":"XiaomiMiMo/MiMo-Embodied repository metadata","date":"2026-06-11T03:20:36.205897+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Embodied","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Embodied\n\nDescription: MiMo-Embodied\n\nLanguage: Python\n\nLicense: NOASSERTION\n\nStars: 386\n\nForks: 15\n\nOpen issues: 0\n\nCreated: 2025-11-19T08:54:41Z\n\nPushed: 2026-04-15T12:28:08Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n## I. Introduction\n\nThis repository provides the **official evaluation suite of MiMo-Embodied**, designed to support **rigorous** and **reproducible** evaluation for **embodied AI** and **autonomous driving** tasks.\n\nBuilt on top of the excellent [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) framework, this repository extends the evaluation pipeline with MiMo-specific model integration, benchmark support, and evaluation workflows for embodied and driving scenarios.\n\n**MiMo-Embodied** is a powerful cross-embodied vision-language model that demonstrates state-of-the-art performance in both **autonomous driving** and **embodied AI tasks**, representing the first open-source VLM that integrates these two critical areas.\n\n> This repository is for **evaluation only**. It does **not** contain model training code.\n\n
\n\n
\n\n---\n\n## II. Key Features\n\n### 1. `MiVLLM`: A MiMo-tailored vLLM-based Model Wrapper\n\nWe use a custom `mivllm` model class built on top of the original `VLLM` implementation in `lmms-eval`, tailored for MiMo models. Compared with the default implementation, it:\n\n- improves **data loading efficiency**\n- enables finer control over **image and video preprocessing**\n- supports MiMo-specific inference settings such as:\n- `max_model_len`\n- `gpu_memory_utilization`\n- `max_num_seqs`\n\n### 2. Evaluation for Embodied AI\n\nThis evaluation suite supports embodied AI benchmarks covering key capabilities such as:\n\n- *"},{"ref":"P9","kind":"page","title":"XiaomiMiMo/MiMo-V2-Flash repository metadata","date":"2026-06-11T03:20:36.044718+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-V2-Flash","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-V2-Flash\n\nDescription: MiMo-V2-Flash: Efficient Reasoning, Coding, and Agentic Foundation Model\n\nLicense: Apache-2.0\n\nStars: 1333\n\nForks: 59\n\nOpen issues: 18\n\nCreated: 2025-12-15T16:28:22Z\n\nPushed: 2026-01-08T04:53:37Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n \n\n# MiMo-V2-Flash\n\n**MiMo-V2-Flash** is a Mixture-of-Experts (MoE) language model with **309B total parameters** and **15B active parameters**. Designed for high-speed reasoning and agentic workflows, it utilizes a novel hybrid attention architecture and Multi-Token Prediction (MTP) to achieve state-of-the-art performance while significantly reducing inference costs.\n\n
\n\n
\n\n-----\n\n## 1. Introduction\n\nMiMo-V2-Flash creates a new balance between long-context modeling capability and inference efficiency. Key features include:\n\n* **Hybrid Attention Architecture**: Interleaves Sliding Window Attention (SWA) and Global Attention (GA) with a 5:1 ratio and an aggressive 128-token window. This reduces KV-cache storage by nearly 6x while maintaining long-context performance via learnable **attention sink bias"},{"ref":"P10","kind":"page","title":"XiaomiMiMo/vllm repository metadata","date":"2026-06-11T02:54:14.938034+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/vllm","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/vllm\n\nDescription: A high-throughput and memory-efficient inference and serving engine for LLMs\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 31\n\nForks: 5\n\nOpen issues: 0\n\nCreated: 2025-04-29T06:43:27Z\n\nPushed: 2025-05-12T09:06:32Z\n\nDefault branch: feat_mimo_mtp_stable_073\n\nFork: yes\n\nParent repository: vllm-project/vllm\n\nArchived: no\n\nREADME:\n
\n\n\n\n\n
\n\n
\nEasy, fast, and cheap LLM serving for everyone\n
\n\n---\n\nWe are excited to invite you to our Menlo Park meetup with Meta, evening of Thursday, February 27! Meta engineers will discuss the improvements on top of vLLM, and vLLM contributors will share updates from the v0.7.x series of releases. [Register Now](https://lu.ma/h7g3kuj9)\n\n---\n\n*Latest News* 🔥\n\n- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).\n- [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).\n- [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving f"},{"ref":"P11","kind":"page","title":"XiaomiMiMo/lmms-eval repository metadata","date":"2026-06-11T02:54:11.272032+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/lmms-eval","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/lmms-eval\n\nDescription: Accelerating the development of large multimodal models (LMMs) with one-click evaluation module - lmms-eval.\n\nLanguage: Python\n\nLicense: NOASSERTION\n\nStars: 71\n\nForks: 5\n\nOpen issues: 0\n\nCreated: 2025-05-29T22:25:08Z\n\nPushed: 2025-08-08T12:15:22Z\n\nDefault branch: mimo_vl_eval\n\nFork: yes\n\nParent repository: EvolvingLMMs-Lab/lmms-eval\n\nArchived: no\n\nREADME:\n# The Evaluation Suite of Xiaomi MiMo-VL\n\nTo promote **rigorous**, **reproducible**, and **thinking-oriented** evaluation of Vision-Language Models (VLMs), we open-source our evaluation suite for [**MiMo-VL**](https://github.com/XiaomiMiMo/MiMo-VL) and beyond.\n\nBuilt on top of the excellent [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) framework, we introduce several improvements in model integration, evaluation protocol, and task coverage to better support the next generation of reasoning-capable VLMs.\n\n## 📰 News\n\n**[25/08/08]** We update our evaluation framework along with the release of [MiMo-VL-7B-SFT-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-SFT-2508) and [MiMo-VL-7B-RL-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508). New features include: \n- Additional GUI action benchmarks AndroidControl and CAGUI (evaluated using `--model mimo_agent`)\n- Additional evaluation benchmarks on video spatial reasoning (VSI-Bench), physics reasoning (PhysReason), multi-modal long context understanding (MMLongBench), multi-modal instruction following (MM-IFEval); \n- Enables no_think evaluation by adding model argument `disable_thinking_user=True`\n\n## 🔧 Key Features\n\n### 1. ⚙️ `MiVLLM`: A vLLM-based Model Wrapper for MiMo-VL\nWe introduce a new `MiVLLM` model class based on the original `VLLM` class in `lmms-eval`, which is tailored for [**MiMo-VL**](https://github.com/XiaomiMiMo/MiMo-VL). Compared to the original implementation, it:\n* Greatly improves **data loading efficiency**\n* Enables fine-grained control over **image and video preprocessing**\n\n### 2. 🧠 Adaptation to Thinking VLMs\n\nThe original lmms-eval tasks were designed for **non-thinking** VLMs: they prompt directly for short answers and compare outputs without post-processing. We redesign this pr"},{"ref":"P12","kind":"page","title":"XiaomiMiMo/MiMo-Skills repository metadata","date":"2026-06-11T02:51:35.366594+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-Skills","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-Skills\n\nDescription: Agent skills for Xiaomi MiMo series.\n\nLanguage: Python\n\nLicense: MIT\n\nStars: 70\n\nForks: 6\n\nOpen issues: 7\n\nCreated: 2026-04-23T12:10:40Z\n\nPushed: 2026-04-24T03:16:07Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n\n\n\n
\n\n> Agent skills for Xiaomi MiMo series. Get your API key at 🎨 Xiaomi MiMo Open Platform.\n\n[中文版](./README_zh.md)\n\n## Skills\n\n| Skill | Description | License |\n|-------|-------------|---------|\n| `mimo-v2-5-tts` | MiMo V2.5 TTS voice synthesis. Supports preset voices, voice design, voice cloning, emotion styles, dialects, and singing. Supports natural language control, Director Mode, and audio tag control. | MIT |\n\n## Installation\n\n### npx (Recommended)\n\n```bash\nnpx skills add XiaomiMiMo/MiMo-Skills\n```\n\nOptions:\n\n```bash\n# List available skills without installing\nnpx skills add XiaomiMiMo/MiMo-Skills --list\n\n# Install to specific agents\nnpx skills add XiaomiMiMo/MiMo-Skills -a claude-code -a opencode\n\n# Install globally (user-level)\nnpx skills add XiaomiMiMo/MiMo-Skills -g\n\n# Install specific skills by name\nnpx skills add XiaomiMiMo/MiMo-Skills --skill mimo-v2-5-tts\n\n# Non-interactive (CI/CD friendly)\nnpx skills add XiaomiMiMo/MiMo-Skills -g -a claude-code -y\n```\n\n### Manual\n\n```bash\ngit clone https://github.com/XiaomiMiMo/MiMo-Skills.git ~/.mimo-skills\nln -s ~/.mimo-skills/skills/* ~/.hermes/skills/\n# ln -s ~/.mimo-skills/skills/* ~/.openclaw/skills/\n# ln -s ~/.mimo-skills/skills/* ~/.config/opencode/skills/\n# ln -s ~/.mimo-skills/skills/* ~/.claude/skills/\n# ...\n```\n\n## Environment Variables\n\n| Variable | Description | Required |\n|----------|-------------|----------|\n| `MIMO_API_KEY` | MiMo API key | Yes |\n| `FEISHU_APP_ID` | Feishu App ID (for voice message sending) | Optional |\n| `FEISHU_APP_SECRET` | Feishu App Secret (for voice messa"},{"ref":"P13","kind":"page","title":"XiaomiMiMo/MiMo-V2.5-ASR repository metadata","date":"2026-06-11T02:51:35.167304+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo-V2.5-ASR","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo-V2.5-ASR\n\nDescription: Robust Speech Recognition Across Languages, Dialects, and Complex Acoustic Scenarios\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 264\n\nForks: 24\n\nOpen issues: 7\n\nCreated: 2026-04-23T12:59:55Z\n\nPushed: 2026-04-23T17:15:38Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n
\n\n
\n\n
\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ \nMiMo-V2.5-ASR: Robust Speech Recognition Across \nLanguages, Dialects, and Complex Acoustic Scenarios \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n\n
\n\n \n\n## Introduction\n\n**MiMo-V2.5-ASR** is a state-of-the-art end-to-end automatic speech recognition (ASR) model developed by the Xiaomi MiMo team. It is built to deliver accurate and robust transcription across Mandarin Chinese and English, multiple Chinese dialects, code-switched speech, song lyrics, knowledge-intensive content, noisy acoustic environments, and multi-speaker conversations. MiMo-V2.5-ASR achieves state-of-the-art results on a wide range of public benchmarks.\n\n## Abstract\n\nAutomatic speech recognition systems are expected to faithfully transcribe speech signals that originate from diverse languages, dialects, accents, and domains, and that are captured under a wide variety of acoustic conditions. While conventional end-to-end models perform well on in-domain data, they still fall short of real-world requirements in challenging scenarios such as dialect mixing, code-switching, knowledge-intensive content, noisy environments, and multi-speaker conversations. Therefore, we present **MiMo-V2.5-ASR**, an end-to-end speech recognition model developed by the Xiaomi MiMo team. Through large-scale mid-training, high-quality supervised fine-"},{"ref":"P14","kind":"page","title":"XiaomiMiMo/MiMo-7B-SFT model card","date":"2026-06-11T02:50:43.687643+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-7B-SFT/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\nlibrary_name: transformers\n---\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n---\n\n## Updates\n\n[2025.05.30] We scaled the SFT dataset from approximately 500K to 6M instances and continuously expanding the RL training window size from 32K to 48K, the performance of [MiMo-7B-RL-0530](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530) on AIME24 can be continuously improved and eventually surpass that of DeepSeek R1 (79.8).\n\n
\n\n
\n
Benchmark
\n
MiMo-7B-RL
\n
MiMo-7B-RL-0530
\n
\n\n\n
\n
Mathematics
\n
\n
\n\n
\n
MATH500 (Pass@1)
95.8
97.2
\n
AIME 2024 (Pass@1)
68.2
80.1
\n
AIME 2025 (Pass@1)
55.4
70.2
\n
Code
\n
LiveCodeBench v5 (Pass@1)
57.8
60.9
\n
LiveCodeBench v6 (Pass@1)
49.3
52.2
\n
STEM
\n
GPQA-Diamond (Pass@1)
54.4
60.6
\n
General
\n\n\n\n\n\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n---\n\n## Updates\n\n[2025.05.30] We scaled the SFT dataset from approximately 500K to 6M instances and continuously expanding the RL training window size from 32K to 48K, the performance of [MiMo-7B-RL-0530](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530) on AIME24 can be continuously improved and eventually surpass that of DeepSeek R1 (79.8).\n\n
\n\n
\n
Benchmark
\n
MiMo-7B-RL
\n
MiMo-7B-RL-0530
\n
\n\n\n
\n
Mathematics
\n
\n
\n\n
\n
MATH500 (Pass@1)
95.8
97.2
\n
AIME 2024 (Pass@1)
68.2
80.1
\n
AIME 2025 (Pass@1)
55.4
70.2
\n
Code
\n
LiveCodeBench v5 (Pass@1)
57.8
60.9
\n
LiveCodeBench v6 (Pass@1)
49.3
52.2
\n
STEM
\n
GPQA-Diamond (Pass@1)
54.4
60.6
\n
General
\n\n\n\n\n\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n---\n\n## Updates\n\n[2025.05.30] We scaled the SFT dataset from approximately 500K to 6M instances and continuously expanding the RL training window size from 32K to 48K, the performance of [MiMo-7B-RL-0530](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530) on AIME24 can be continuously improved and eventually surpass that of DeepSeek R1 (79.8).\n\n
\n\n
\n
Benchmark
\n
MiMo-7B-RL
\n
MiMo-7B-RL-0530
\n
\n\n\n
\n
Mathematics
\n
\n
\n\n
\n
MATH500 (Pass@1)
95.8
97.2
\n
AIME 2024 (Pass@1)
68.2
80.1
\n
AIME 2025 (Pass@1)
55.4
70.2
\n
Code
\n
LiveCodeBench v5 (Pass@1)
57.8
60.9
\n
LiveCodeBench v6 (Pass@1)
49.3
52.2
\n
STEM
\n
GPQA-Diamond (Pass@1)
54.4
60.6
\n
General
\n\n\n\n\n\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n> This model repository is licensed under the MIT License.\n\n## I. Introduction\n\nCurrently, most successful RL works, including open-source research, rely on relatively large base models, e.g., 32B models, particularly for enhancing code reasoning capabilities. Moreover, it was widely considered that achieving uniform and simultaneous improvements in both mathematical and code capabilities within a small model is challenging. Nonetheless, we believe that the effectiveness of the RL trained reasoning model relies on the inherent reasoning potential of the base model. To fully unlock the reasoning potential of language models, efforts must focus not only on post-training but also on pre-training strategies tailored to reasoning.\n\nIn this work, we present MiMo-7B, a series of models trained from scratch and born for reasoning tasks. Our RL experiments from MiMo-7B-Base show that our model possesses extraordinary reasoning potential, even surpassing much larger 32B models. Additionally, we perform RL training on a cold-started SFT model, resulting in MiMo-7B-RL, which demonstrates superior performance on both mathematics and code reasoning tasks, matching the performance of OpenAI o1-mini.\n\n
\n\n \n\n## I. Introduction\n\nIn this report, we share our efforts to build a compact yet powerful VLM, MiMo-VL-7B. MiMo-VL-7B comprises (1) a native resolution ViT encoder that preserves fine-grained visual details, (2) an MLP projector for efficient cross-modal alignment, and (3) our [MiMo-7B language model](https://github.com/XiaomiMiMo/MiMo), specifically optimized for complex reasoning tasks. \n\nThe development of MiMo-VL-7B involves two sequential training processes: (1) A four-stage pre-training phase, which includes projector warmup, vision-language alignment, general multi-modal pre-training, and long-context Supervised Fine-Tuning (SFT). This phase yields the MiMo-VL-7B-SFT model. (2) A subsequent post-training phase, where we introduce Mixed On-policy Reinforcement Learning (MORL), a novel framework that seamlessly integrates diverse reward signals spanning perception accuracy, visual grounding precision, logical reasoning capabilities, and human/AI preferences. This phase yields the MiMo-VL-7B-RL model.\n\n
\n\n\n\n\n\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n---\n\n## Updates\n\n[2025.05.30] We scaled the SFT dataset from approximately 500K to 6M instances and continuously expanding the RL training window size from 32K to 48K, the performance of [MiMo-7B-RL-0530](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-0530) on AIME24 can be continuously improved and eventually surpass that of DeepSeek R1 (79.8).\n\n
\n\n \n\n## I. Introduction\n\nIn this report, we share our efforts to build a compact yet powerful VLM, MiMo-VL-7B. MiMo-VL-7B comprises (1) a native resolution ViT encoder that preserves fine-grained visual details, (2) an MLP projector for efficient cross-modal alignment, and (3) our [MiMo-7B language model](https://github.com/XiaomiMiMo/MiMo), specifically optimized for complex reasoning tasks. \n\nThe development of MiMo-VL-7B involves two sequential training processes: (1) A four-stage pre-training phase, which includes projector warmup, vision-language alignment, general multi-modal pre-training, and long-context Supervised Fine-Tuning (SFT). This phase yields the MiMo-VL-7B-SFT model. (2) A subsequent post-training phase, where we introduce Mixed On-policy Reinforcement Learning (MORL), a novel framework that seamlessly integrates diverse reward signals spanning perception accuracy, visual grounding precision, logical reasoning capabilities, and human/AI preferences. This phase yields the MiMo-VL-7B-RL model.\n\n
\n\n \n\n### 🔥🔥🔥MiMo-VL 2508 Updates\n\nWe're excited to announce improvements to our MiMo-VL ([MiMo-VL-7B-RL-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508) and [MiMo-VL-7B-SFT-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-SFT-2508)), featuring enhanced performance across multiple benchmarks, improved thinking control capabilities, and better user experience.\n\n#### 📈 Performance Improvements\n\nMiMo-VL-7B-RL-2508 demonstrates consistent improvements across both image and video benchmarks, achieving notable milestones of **70.6 on MMMU** and **70.8 on VideoMME**.\n\n\n\nFull evaluation results can be found [below](#full-evaluation-results).\n\n#### 🤔 Thinking Control Feature\n\nA thinking control capability that allows users to turn off the model's reasoning mode using the no_think parameter:\n- Thinking mode (default behavior): Full reasoning process visible with 100% control success rate;\n- Non-thinking mode: Direct responses without reas"},{"ref":"P22","kind":"page","title":"XiaomiMiMo/MiMo-VL-7B-RL-2508 model card","date":"2026-06-11T02:50:37.928399+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nbase_model:\n- XiaomiMiMo/MiMo-VL-7B-RL-2508\nlibrary_name: transformers\nlicense: mit\npipeline_tag: image-text-to-text\n---\n\n
\n\n \n\n### 🔥🔥🔥MiMo-VL 2508 Updates\n\nWe're excited to announce improvements to our MiMo-VL ([MiMo-VL-7B-RL-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508) and [MiMo-VL-7B-SFT-2508](https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-SFT-2508)), featuring enhanced performance across multiple benchmarks, improved thinking control capabilities, and better user experience.\n\n#### 📈 Performance Improvements\n\nMiMo-VL-7B-RL-2508 demonstrates consistent improvements across both image and video benchmarks, achieving notable milestones of **70.6 on MMMU** and **70.8 on VideoMME**.\n\n\n\nFull evaluation results can be found [below](#full-evaluation-results).\n\n#### 🤔 Thinking Control Feature\n\nA thinking control capability that allows users to turn off the model's reasoning mode using the no_think parameter:\n- Thinking mode (default behavior): Full reasoning process visible with 100% control success rate;\n- Non-thinking mode: Direct responses without reaso"},{"ref":"P23","kind":"page","title":"XiaomiMiMo/MiMo-Audio-7B-Base model card","date":"2026-06-11T02:50:34.916209+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-Audio-7B-Base/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\npipeline_tag: any-to-any\ntags:\n- Audio-to-Text\n- Text-to-Audio\n- Audio-to-Audio\n- Text-to-Text\n- Audio-Text-to-Text\n---\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nMiMo Audio: Audio Language Models are Few-Shot Learners\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n## Introduction\n\nExisting audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversi"},{"ref":"P24","kind":"page","title":"XiaomiMiMo/MiMo-Audio-Tokenizer model card","date":"2026-06-11T02:50:34.779809+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-Audio-Tokenizer/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\n---\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nMiMo Audio: Audio Language Models are Few-Shot Learners\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n## Introduction\n\nExisting audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demons"},{"ref":"P25","kind":"page","title":"XiaomiMiMo/MiMo-Audio-7B-Instruct model card","date":"2026-06-11T02:50:34.568088+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-Audio-7B-Instruct/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\npipeline_tag: any-to-any\ntags:\n- Audio-to-Text\n- Text-to-Audio\n- Audio-to-Audio\n- Text-to-Text\n- Audio-Text-to-Text\n---\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nMiMo Audio: Audio Language Models are Few-Shot Learners\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n## Introduction\n\nExisting audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversi"},{"ref":"P26","kind":"page","title":"XiaomiMiMo/MiMo-7B-MTPs model card","date":"2026-06-11T02:49:53.014243+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-7B-MTPs/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\nlibrary_name: transformers\n---\n\n
\n\n\n\n\n
\n\n
\n\n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \nUnlocking the Reasoning Potential of Language Model From Pretraining to Posttraining\n \n━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\n \n\n
\n\n \n\n> This model repository is licensed under the MIT License.\n\n## I. Pretrained MTPs of MiMo-7B\n\nThis model repository contains the pretrained MTP weights of MiMo-7B (`model.mtp_layers.1` and `model.mtp_layers.2`)\n\nCurrently, MiMo-7B model each has 1 MTP layer (`model.mtp_layers.0`). Users may load the weights of pretrained MTPs for potential rollout speedup (please refer to *[Power Up Speculative Decoding In Reinforcement Learning](https://www.notion.so/jiajunli-guapisolo/Power-Up-Speculative-Decoding-In-Reinforcement-Learning-2a92d24a293b802d9c73dbae429e581e)*).\n\n> [!IMPORTANT]\n> We tuned 1 MTP layer in SFT and freeze it in RL, and we **HAVE NOT** test the performance of posttrained models with 2 more pretrained MTP layers.\n\n## II. Contact\n\nPlease contact us at [mimo@xiaomi.com](mailto:mimo@xiaomi.com) or open an issue if you have any questions."},{"ref":"P27","kind":"page","title":"XiaomiMiMo/MiMo-Embodied-7B model card","date":"2026-06-11T02:49:53.008774+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-Embodied-7B/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nbase_model:\n- XiaomiMiMo/MiMo-Embodied\nlibrary_name: transformers\nlicense: mit\n---\n\n
\n\n## I. Introduction\n\n**MiMo-Embodied**, a powerful cross-embodied vision-language model that shows state-of-the-art performance in both **autonomous driving** and **embodied AI tasks**, the first open-source VLM that integrates these two critical areas, significantly enhancing understanding and reasoning in dynamic physical environments.\n\n
\n\n
\n\n## II. Model Capabilities\n\n
\n\n
\n\n## III. Model Details\n\n
\n\n
\n\n## IV. Evaluation Results\n\nMiMo-Embodied demonstrates superior performance across **17 benchmarks in three key embodied AI capabilities: Task Planning, Affordance Prediction, and Spatial Understanding**, significantly surpassing existing open-source embodied VLM models and rivaling closed-source models.\n\nAdditionally, MiMo-Embodied excels in **12 autonomous driving benchmarks across three key capabilities: Environmental Perception, Status Prediction, and Driving Planning**—significantly outperforming both existing open-source and closed-source VLM models, as well as proprietary VLM models.\n\nMoreover, evaluation on **8 general visual understanding benchmarks** confirms that MiMo-Embodied retains and even strengthens its general capabilities, showing that domain-specialized training enhances rather than diminishes overall model proficiency.\n\n### Embodied AI Benchmarks\n\n#### Affordance & Planning\n\n
"},{"ref":"P28","kind":"page","title":"XiaomiMiMo/MiMo-V2-Flash model card","date":"2026-06-11T02:49:48.056976+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nlicense: mit\nlibrary_name: transformers\n---\n\n
\n \n\n# MiMo-V2-Flash\n\n**MiMo-V2-Flash** is a Mixture-of-Experts (MoE) language model with **309B total parameters** and **15B active parameters**. Designed for high-speed reasoning and agentic workflows, it utilizes a novel hybrid attention architecture and Multi-Token Prediction (MTP) to achieve state-of-the-art performance while significantly reducing inference costs.\n\n
![\"Xiaomi-MiMo\"](\"https://github.com/XiaomiMiMo/MiMo-VL/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true\")
\n![\"Benchmark](\"./figures/mimo-2508.png\")
\n\nFull evaluation results can be found [below](#full-evaluation-results).\n\n#### 🤔 Thinking Control Feature\n\nA thinking control capability that allows users to turn off the model's reasoning mode using the no_think parameter:\n- Thinki"},{"ref":"P2","kind":"page","title":"XiaomiMiMo/MiMo repository metadata","date":"2026-06-11T03:20:37.087985+00:00","date_source":null,"source_url":"https://github.com/XiaomiMiMo/MiMo","signal_url":null,"signal_json_url":null,"text":"# XiaomiMiMo/MiMo\n\nDescription: MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining\n\nLanguage: Python\n\nLicense: Apache-2.0\n\nStars: 2167\n\nForks: 94\n\nOpen issues: 38\n\nCreated: 2025-04-26T09:31:17Z\n\nPushed: 2025-06-05T16:01:49Z\n\nDefault branch: main\n\nFork: no\n\nArchived: no\n\nREADME:\n![\"Xiaomi-MiMo\"](\"https://github.com/XiaomiMiMo/MiMo/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true\")
\n![](\"https://github.com/XiaomiMiMo/MiMo/raw/main/figures/length.jpg?raw=true\")
![\"Description\"](\"https://raw.githubusercontent.com/XiaomiMiMo/MiMo-VL/main/figures/Xiaomi_MiMo.png\")
\n\n# MiMo-Audio-Tokenizer\n\n![\"Description\"](\"mimo_audio_tokenizer/assets/Tokenizer_01.png\")
\n\n![](\"./assets/xfmlogo.svg\")
\n![](\"./assets/fig1.svg\")
\n![\"Xiaomi-MiMo\"](\"https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true\")
\n![](\"https://github.com/XiaomiMiMo/MiMo-V2-Flash/raw/main/figures/MiMo-v2-flash-performance.jpg?raw=true\")
\n![\"vLLM\"](\"https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-dark.png\"){kind=logo}
\n![\"Xiaomi-MiMo\"](\"assets/XiaomiMIMO.png\")
\n![\"Benchmark](\"./mimo-2508.png\")
\n\nFull evaluation results can be found [below](#full-evaluation-results).\n\n#### 🤔 Thinking Control Feature\n\nA thinking control capability that allows users to turn off the model's reasoning mode using the no_think parameter:\n- Thinking mode (default behavior): Full reasoning process visible with 100% control success rate;\n- Non-thinking mode: Direct responses without reas"},{"ref":"P22","kind":"page","title":"XiaomiMiMo/MiMo-VL-7B-RL-2508 model card","date":"2026-06-11T02:50:37.928399+00:00","date_source":null,"source_url":"https://huggingface.co/XiaomiMiMo/MiMo-VL-7B-RL-2508/raw/main/README.md","signal_url":null,"signal_json_url":null,"text":"---\nbase_model:\n- XiaomiMiMo/MiMo-VL-7B-RL-2508\nlibrary_name: transformers\nlicense: mit\npipeline_tag: image-text-to-text\n---\n\n![](\"./assets/fig2.svg\")
\n![](\"./assets/fig3_img.png\")
\n![](\"./assets/table2.png\")
\n![](\"./assets/table3.png\")
\n