Source-linked AI summary
MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining
LLM-Core Xiaomi, :, Bingquan Xia, Bowen Shen, Cici, Dawei Zhu, Di Zhang, Gang Wang, Hailin Zhang, Huaqiu Liu, Jiebao Xiao, Jinhao Dong, Liang Zhao, Peidian Li, Peng Wang, Shihua Yu, Shimao Chen, Weikun Wang, Wenhan Ma, Xiangwei Deng, Yi Huang, Yifan Song, Zihan Jiang, Bowen Ye, Can Cai, Chenhong He, Dong Zhang, Duo Zhang, Guoan Wang, Hao Tian, Haochen Zhao, Heng Qu, Hongshen Xu, Jun Shi, Kainan Bao, Kai Fang, Kang Zhou, Kangyang Zhou, Lei Li, Menghang Zhu, Nuo Chen, Qiantong Wang, Shaohui Liu, Shicheng Li, Shuhao Gu, Shuhuai Ren, Shuo Liu, Sirui Deng, Weiji Zhuang, Weiwei Lv, Wenyu Yang, Xin Zhang, Xing Yong, Xing Zhang, Xingchen Song, Xinzhe Xu, Xu Wang, Yihan Yan, Yu Tu, Yuanyuan Tian, Yudong Wang, Yue Yu, Zhenru Lin, Zhichao Song, Zihao Yue
TL;DR
Small models were widely considered unlikely to improve mathematics and code reasoning uniformly, motivating attention to both pre-training and post-training. MiMo-7B combines reasoning-focused model development with reinforcement-learning strategies, and its models achieve strong mathematics, code, and general reasoning performance, including results surpassing OpenAI o1-mini.
Problem
Small models were widely considered unlikely to achieve uniform improvements in both mathematical and code reasoning, while successful reinforcement-learning work often relied on larger base models.
Method
MiMo-7B optimizes pre-training and post-training, using reasoning-focused model development, test-difficulty-driven rewards, and easy-data resampling for reinforcement learning.
Results
MiMo-7B models demonstrate superior reasoning across mathematics, code, and general tasks, with MiMo-7B-RL scoring 55.4 on AIME 2025 and exceeding o1-mini by 4.7 points.
Takeaways & Limitations
MiMo-7B shows that a 7B model can develop strong reasoning potential through coordinated pre-training and post-training, including performance exceeding much larger models.
Takeaways & Limitations
A language-mixing penalty does not fully resolve language mixing and introduces a risk of reward hacking, such as always generating English responses.
Abstract
from arXiv · showhide
We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.
1 Introduction
MiMo-7B targets the challenge of unlocking reasoning in small models through coordinated pre-training and post-training. Its base and RL-tuned variants report strong results across general reasoning, mathematics, and code.
- MiMo-7B addresses the view that strong RL reasoning typically depends on 32B models and that small models struggle to improve mathematics and code simultaneously.
- Pre-Training: Base Model Born for Reasoning: The pre-training recipe combines improved data preprocessing, three-stage data mixing, and Multi-Token Prediction to strengthen reasoning potential and accelerate inference.
- Post-Training Recipe: Pioneering Reasoning Model: The post-training recipe uses 130K verifiable mathematics and code problems, test-difficulty-driven code rewards, and data resampling for denser rewards and more stable policy updates.
- 75.2 on BBH shows MiMo-7B-Base outperforming comparable open-source models in general reasoning, knowledge, and coding evaluations.
- 55.4 on AIME 2025 exceeds o1-mini by 4.7 points, while MiMo-7B-RL significantly outperforms it on LiveCodeBench v5 and v6.
- MiMo-7B checkpoints, including base, SFT, and RL variants, are released openly for further development of reasoning language models.
2 Pre-Training
MiMo-7B-Base is designed to acquire dense reasoning patterns during large-scale pre-training while improving both reasoning performance and inference efficiency. Evaluations report strong results across pass@k, general reasoning, coding, mathematics, and long-context comprehension.
- Three-Stage Data Mixture: The second mixture stage raises mathematics and code data to approximately 70%, while the final stage adds approximately 10% synthetic responses and expands context from 8,192 to 32,768 tokens.
- Pre-Training Data: MiMo-7B-Base is trained on approximately 25 trillion tokens using enhanced extraction, filtering, synthetic reasoning data, and a three-stage mixture strategy.
- MTP Modules: Multi-Token Prediction addresses autoregressive inference bottlenecks by preparing representations for future tokens; pre-training uses one MTP layer, while inference can use multiple.
- Evaluation Results: MiMo-7B-Base achieves significantly higher pass@k scores than compared models, including the 32B baseline, across benchmarks and evaluated k values.
- Evaluation Results: 75.2 on BBH and 32.9 on both LiveCodeBench v5 and AIME 2024 demonstrate strong general reasoning, coding, and mathematics performance.
- Long-Context Comprehension: MiMo-7B achieves near-perfect NIAH retrieval across positions within the 32K context window and surpasses Qwen2.5-7B in most reported long-context reasoning scenarios.
3 Post-Training
Post-training proceeds from MiMo-7B-Base through direct RL and from an SFT version of MiMo-7B. The SFT pipeline filters data for quality, diversity, and benchmark decontamination before fine-tuning.
- MiMo-7B-RL-Zero is trained by direct RL from MiMo-7B-Base, while MiMo-7B-RL is trained from an SFT version of MiMo-7B.
- SFT Data: The SFT data combines open-source and proprietary distilled data processed through a three-stage quality and diversity pipeline.
- SFT Data: Training-query filtering removes samples with 16-gram overlap with evaluation benchmarks to prevent data leakage, alongside exclusion of mixed-language or incomplete responses.
- SFT Data: The final SFT dataset contains about 500K samples after capping responses per query at eight to balance diversity and redundancy.
- SFT Hyper-parameters: SFT fine-tuning uses a constant learning rate of 3 × 10^-5, batch size 128, and samples packed to a maximum length of 32,768 tokens.
3.2 RL Data Curation
MiMo-7B forms its RL dataset from verifiable mathematics and code problems, emphasizing quality control, difficulty assessment, and rule-based rewards. The resulting curation supports stable and efficient RL training.
- RL training data uses two verifiable categories—mathematics and code—because high-quality problem sets help stabilize training and enhance reasoning capabilities.
- Mathematical Data: Mathematical problems are cleaned through filtering, preservation of original formulations, global n-gram deduplication, and decontamination to reduce reward hacking.
- Mathematical Data: Difficulty assessment removes unsolvable or incorrect problems and excludes problems with pass rates above 90%, eliminating approximately 50% of easy problems.
- Code Data: Code curation removes problems without tests, invalid golden solutions, or insufficiently solvable test cases under advanced-model rollouts.
- Reward Function: Rule-based accuracy rewards use Math-Verify for mathematics and test-based rewards for code, without format or length-penalty rewards.
3.3 RL Training Recipe
MiMo’s RL recipe modifies GRPO with dynamic sampling and Clip-Higher, while addressing sparse code rewards and declining sampling efficiency through difficulty-aware rewards and easy-data resampling.
- GRPO Enhancements: MiMo uses a modified GRPO objective that samples response groups from the old policy and updates the current policy using group-relative advantages.The recipe also incorporates removal of KL loss, dynamic sampling, and Clip-Higher.
- GRPO Enhancements: Dynamic sampling filters prompts with pass rates of 0 or 1, keeping effective gradients while maintaining a consistent batch size and calibrating difficulty during training.Clip-Higher raises the upper clipping bound while fixing the lower bound to encourage exploration and mitigate entropy convergence.
- Reward Design: Difficult code problems can produce no reward under all-tests-passed evaluation, preventing learning from challenging cases and reducing dynamic-sampling efficiency.This sparse-reward issue motivates a test difficulty driven code reward.
- Reward Design: Test difficulty is estimated by rolling out several models and clustering test cases by pass rate, with lower pass rates indicating greater difficulty.The method is inspired by IOI-style partial scoring, where subtasks receive points separately.
- Reward Design: The strict reward requires passing all tests in a difficulty group and easier groups, whereas the soft reward sums equally distributed scores for passed tests.Figure 5 compares both schemes with a baseline without test difficulty driven reward.
- Data Resampling: Easy data resampling stores perfectly solved problems in an easy pool and samples from it with probability α=10%, improving efficiency while stabilizing policy updates.Removing perfect-pass problems entirely caused significant instability in preliminary studies.
3.4 RL Infrastructures
The Seamless Rollout Engine improves RL infrastructure by coordinating rollout, reward computation, and termination asynchronously to reduce GPU idle time and accelerate training and validation.
- Seamless Rollout Engine: The Seamless Rollout Engine integrates continuous rollout, asynchronous reward computation, and early termination, achieving 2.29× faster training and 1.96× faster validation.It is built on verl and extends vLLM for MiMo’s inference requirements.
- Engine Components: Continuous rollout removes synchronization barriers by rewarding completed workers immediately and launching new rollouts on demand.The engine monitors valid-sample counts and pass-rate statistics to decide when additional tasks are needed.
- Engine Components: Asynchronous reward computation concurrently manages rollout and reward tasks, addressing the high overhead of code judging and underused multiprocessing resources.Ray dynamically forwards completed rollout outputs for evaluation or aggregates results for sample updates.
- Engine Components: Early termination reduces waiting after enough valid samples are available while avoiding abrupt cancellation that could suppress long-sequence responses and destabilize training.The design balances latency against the training dynamics of long responses.
- Experimental Analysis: All three components contribute to faster dynamic sampling and lower GPU idle time, whereas static sampling produces many zero-gradient samples that reduce effective batch size and risk instability.The comparison used a randomly selected five-step trace on 256 H20 GPUs.
- Accelerated Validation: Validation reaches 1.96× speedup while reducing idle GPU time to 25% with asynchronous reward computation.The experiment used the full validation dataset on 256 H20 GPUs.
3.5 Post-Training Evaluation
MiMo-7B is evaluated across language understanding, science, instruction following, reading comprehension, mathematics, and coding benchmarks against both non-reasoning and reasoning-model baselines.
- Benchmarks: The evaluation covers MMLU-Pro, GPQA Diamond, SuperGPQA, IFEval, DROP, MATH500, AIME 2024, AIME 2025, and LiveCodeBench v5 and v6.Repeated sampling is used for several benchmarks, including AIME and LiveCodeBench.
- Baselines: MiMo-7B-RL is compared with GPT-4o-0513, Claude-Sonnet-3.5-1022, OpenAI-o1-mini, QwQ-32B-Preview, and DeepSeek-R1 distill models.The comparison includes both non-reasoning and reasoning baselines.
- Evaluation Results: MiMo-7B-RL achieves top-tier mathematics performance among similarly sized models and strong general performance, while excelling on algorithmic code generation.The model is evaluated with temperature 0.6, top-p 0.95, and task-dependent maximum generation lengths.
- Evaluation Results: 49.3% on LiveCodeBench v6 surpasses QwQ-32B-Preview by over 10 points, while LiveCodeBench v5 significantly outperforms OpenAI o1-mini.The authors describe these results as robust and stable capabilities.
- MiMo Variants: RL from the base model shows stronger growth, while RL from the SFT model reaches a higher performance ceiling and achieves the best results across evaluated benchmarks.MiMo-7B-RL-Zero is trained from MiMo-7B-Base, whereas MiMo-7B-RL is trained from MiMo-7B-SFT.
3.6 Discussion
The discussion examines post-training choices that affect reasoning performance, including SFT scaling, domain interference, language-mixing penalties, and extended on-policy RL. It also reports limitations and performance effects observed across model variants and training settings.
- SFT for Format Alignment: The light-weight SFT variant fell behind the base model’s trajectory after 500 steps and ultimately produced poorer final results.Although it began above MiMo-7B-RL-Zero, its inferior starting point limited final performance relative to the heavier-SFT model.
- Interference Between Different Domains: MiMo-7B-Base showed code improvement but mathematical fluctuation and decline between steps 2000 and 2500, unlike the cold-started SFT model.The authors attribute this contrast to reward hacking in mathematics, whereas test-case verification made exploitation harder for code.
- Language Mixing Penalty: The language-mixing penalty did not fully resolve mixed-language outputs and introduced a risk of reward hacking.The reverse detection problem is especially difficult because mathematical equations and code contain English words.
- Impact of SFT Data Scaling: 6M SFT instances improved mathematical, code, scientific, and dialogue abilities while preserving subsequent RL gains.The expanded dataset was compared with approximately 500K instances, and models fine-tuned with RL afterward sustained performance improvements.
- On-Policy RL with Extended Generation Budget: On-policy RL was adopted because vanilla GRPO was prone to premature performance saturation, while the on-policy approach enabled stable, sustained growth.The passage presents this as a training-design response to saturation during post-training.
- On-Policy RL with Extended Generation Budget: Increasing generation length from 32K to 38K and then 48K consistently boosted performance during on-policy RL.The progressive budget extension was reported as instrumental in enabling the 7B model to reach parity with DeepSeek-R1 on mathematical reasoning.
4 Conclusion
The conclusion presents MiMo-7B as a reasoning-focused model developed through optimized pre-training and post-training. Its base and RL-trained variants demonstrate strong reasoning capabilities across mathematics, code, and general tasks.
- 4 Conclusion: MiMo-7B unlocks advanced reasoning capabilities through optimized pre-training and post-training processes.The paper describes the model as exposed to diverse reasoning patterns during pre-training and supported by robust, efficient RL frameworks.
- 4 Conclusion: MiMo-7B-Base possesses exceptional reasoning potential and outperforms models of significantly larger scale.The conclusion attributes this potential to exposure to diverse reasoning patterns during pre-training.
- 4 Conclusion: MiMo-7B-RL-Zero and MiMo-7B-RL demonstrate superior reasoning capabilities across mathematics, code, and general tasks.These variants were trained using the paper’s post-training frameworks.
A Contributions and Acknowledgments
The acknowledgments section expresses gratitude to contributors, including individuals not listed in the paper, and describes the author-listing convention.
- A Contributions and Acknowledgments: The authors thank all contributors, including those not listed in the paper, for their support and efforts.The acknowledgment is explicitly directed to contributors whose work supported the project.
- A Contributions and Acknowledgments: Authors within each role are listed alphabetically by first name.This specifies the ordering convention used for the author list.
Core Contributors
The core-contributors section lists the individuals credited with the work.
- Core Contributors: The listed core contributors are Bingquan Xia, Bowen Shen, Cici Dawei Zhu, Di Zhang, Gang Wang, Hailin Zhang, and Huaqiu Liu.The list continues with Jiebao Xiao, Jinhao Dong, Liang Zhao, Peidian Li, Peng Wang, Shihua Yu, Shimao Chen, Weikun Wang, Wenhan Ma, Xiangwei Deng, Yi Huang, Yifan Song, and Zihan Jiang.