Modelpublic

xiaomi-open-source/Xiaomi-MiMo-VL-Miloco-7B-GGUF

sourceHugging Faceupdated 11mo agoView on Hugging Face

9likes815downloads

Model Card

<br/>

<div align="center" style="line-height: 1;"> <a href="https://huggingface.co/xiaomi-open-source/Xiaomi-MiMo-VL-Miloco-7B" target="blank">🤗 HuggingFace</a>  | <a href="https://modelscope.cn/models/xiaomi-open-source/Xiaomi-MiMo-VL-Miloco-7B" target="blank">🤖️ ModelScope</a> <br/> </div>

<br/>

MiMo-VL-7B-GGUF

This repository provides a quantized version of MiMo-VL-Miloco-7B. It includes two components:

—Language model (LLM): Q4_0
—Visual encoder (mmproj): BF16

These files are built with ``llama.cpp``, support inference on multiple platforms, and are intended for resource-constrained scenarios.

Introduction

Welcome to Xiaomi MiMo-VL-Miloco — the first open-source multimodal model built to actually understand what’s happening at home!

🤗 Why you’ll love it:

—Built on MiMo-VL-7B: a rock-solid vision–language backbone with reliable video understanding and instruction-following.
—Home-savvy by design: it spots everyday activities (esports, workouts, watching TV, reading, and more) and reads common hand gestures like the V sign, thumbs-up, open palm, OK, and even the shaka hand sign.
—Base skills intact: with a mix training strategy of SFT and RL, we boost home-scene smarts while keeping the model’s generality and transferability in great shape.

🌟 Training recipe:

We use a carefully tuned two-stage pipeline to nail home-scene skills without sacrificing general abilities.

Stage 1: Supervised Fine-Tuning (SFT)

This stage focuses on boosting the model’s core capabilities in home scenarios. Even with a limited training set, we strike a good balance between sample-efficient learning and fast inference:

—Chain-of-thought supervision: we add chain of reasoning so the model learns structured knowledge about home scenarios.
—Token-budget-aware reasoning: training with “budgeted” reasoning encourages concise, straight-to-the-point answers at inference.

Stage 2: Reinforcement Learning (RL)

Building on fine-tuning, this stage introduces GRPO-based reinforcement learning to enhance the model’s overall performance:

—Efficient Training Data: we employed the Time-R1 data strategy (our work accepted at NeurIPS 2025) to build efficient training datasets across multiple domains.
—Keep-it-general: specialize for home tasks while preserving broad understanding and language generation.

In short: Xiaomi MiMo-VL-Miloco is your friendly, sharp-eyed model roommate—great at recognizing what’s going on around the house, and still ready for the wider world.

Performance

Evaluation of Home-Scenario Undersatnding Capabilities (F1-Score)

—MiMo-VL-Miloco-7B achieves leading performance in both gesture recognition and common household scene understanding.

Results of general capability evaluations

In household scene understanding, we prioritize video and image perception alongside the model’s reasoning ability.

—Across three video benchmarks (Video-MME, Video-MMMU, Charades-STA), the base model shows clear improvements.
—On MMMU-Pro, a general-capabilities benchmark, the base model also saw significant improvements (10+%).
—Surprisingly, as video and image understanding improved, we observed corresponding gains on the text-only task MMLU-Pro.
—We see a modest performance dip on tasks such as document understanding, OCR, and mathematics; this is in line with expectations and does not affect the model’s intended use cases.

Citation

bibtex

@misc{xiaomimimovlmiloco,
  author       = {Jiaze Li, Yuxun Qu, Jingyang Chen, Shijie Xu, Zhenru Lin, Junyou Zhu, Boshen Xu, Wenhui Tan, Pei Fu, JianZhong Ju, Zhenbo Luo, Jian Luan},
  title        = {Xiaomi MiMo-VL-Miloco},
  year         = {2025},
  howpublished = {\url{https://github.com/XiaoMi/xiaomi-mimo-vl-miloco}},
}

Contact

Please contact us at milm-plus@xiaomi.com or open an issue if you have any questions.