We open-source MiMo-VL-7B-SFT and MiMo-VL-7B-RL, two powerful vision-language models delivering state-of-the-art performance in both general visual understanding and multimodal reasoning. MiMo-VL-7B-RL outperforms Qwen2.5-VL-7B on 35 out of 40 evaluated tasks, and scores 59.4 on OlympiadBench, surpassing models with up to 78B parameters. For GUI grounding applications, it sets a new standard with 56.1 on OSWorld-G, even outperforming specialized models such as UI-TARS. Our training combines four-stage pre-training (2.4 trillion tokens) with Mixed On-policy Reinforcement Learning (MORL) integrating diverse reward signals. We identify the importance of incorporating high-quality reasoning data with long Chain-of-Thought into pre-training stages, and the benefits of mixed RL despite challenges in simultaneous multi-domain optimization. We also contribute a comprehensive evaluation suite covering 50+ tasks to promote reproducibility and advance the field. The model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-VL.
核心贡献 · Key contributions
开源 MiMo-VL-7B-SFT 和 MiMo-VL-7B-RL，在 35/40 项任务上超越 Qwen2.5-VL-7B，达到领先水平。 Open-sources MiMo-VL-7B-SFT and MiMo-VL-7B-RL, achieving state-of-the-art performance on 35/40 tasks vs Qwen2.5-VL-7B.
提出混合在线强化学习（MORL），整合多种奖励信号，实现多领域稳定优化。 Proposes Mixed On-policy Reinforcement Learning (MORL) integrating diverse reward signals for stable multi-domain optimization.
发现将长思维链推理数据纳入预训练阶段对提升性能至关重要。 Identifies importance of incorporating long Chain-of-Thought reasoning data into pre-training stages for enhanced performance.
在 OlympiadBench 上取得 59.4 分，超越高达 78B 参数的模型；在 OSWorld-G 上取得 56.1 分，刷新 GUI 接地标准。 Achieves 59.4 on OlympiadBench, surpassing models with up to 78B parameters, and 56.1 on OSWorld-G for GUI grounding.
贡献包含 50+ 任务的综合评估套件，提供提示词和协议以促进可复现性。 Contributes a comprehensive evaluation suite covering 50+ tasks with prompts and protocols for reproducibility.
在用户偏好评估中，Elo 评分在开源 VLM 中最高，与闭源模型竞争。 Demonstrates highest Elo rating among open-source VLMs in user preference evaluation, competitive with proprietary models.
局限 · Limitations
MORL 面临任务干扰，推理与接地任务响应长度趋势相反。 MORL faces task interference due to opposing response length trends between reasoning and grounding tasks.
标准 GRPO 在约 2 万样本时饱和，而在线 RL 在观察窗口内未饱和。 Vanilla GRPO saturates around 20k samples, while on-policy RL shows no saturation within observed training window.
多领域同步优化仍具挑战；单任务独立 RL 可能优于 MORL。 Simultaneous multi-domain optimization remains challenging; standalone RL on individual tasks may outperform MORL.
评估限于 7B 参数规模；扩展到更大模型可能呈现不同行为。 Evaluation limited to 7B parameter scale; scaling to larger models may reveal different behaviors.
合成推理数据质量依赖大型推理模型；拒绝采样可能无法消除所有错误。 Synthetic reasoning data quality depends on large reasoning models; rejection sampling may not eliminate all errors.
论文章节 · Sections（共 19）
摘要Abstract
1 引言1 Introduction
2 预训练2 Pre-Training
2.1 架构2.1 Architecture
2.2 预训练数据2.2 Pre-training Data
2.3 预训练阶段2.3 Pre-training Stages
3 后训练3 Post-Training
3.1 基于可验证奖励的强化学习3.1 Reinforcement Learning with Verifiable Rewards
3.2 基于人类反馈的强化学习3.2 Reinforcement Learning from Human Feedback