XiaomiMiMo
/

MiMo-7B-RL

@@ -1,7 +1,6 @@
 ---
 license: mit
 ---
 <div align="center">
   <picture>
     <source srcset="https://github.com/XiaomiMiMo/MiMo/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true" media="(prefers-color-scheme: dark)">
@@ -11,11 +10,11 @@ license: mit
 <h3 align="center">
   <b>
-    <span>━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━</span>
     <br/>
     Unlocking the Reasoning Potential of Language Model<br/>From Pretraining to Posttraining
     <br/>
-    <span>━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━</span>
     <br/>
   </b>
 </h3>
@@ -26,6 +25,8 @@ license: mit
   |
   <a href="https://huggingface.co/XiaomiMiMo" target="_blank">🤗 HuggingFace</a>
   &nbsp;|
   <a href="https://github.com/XiaomiMiMo/MiMo/blob/main/MiMo-7B-Technical-Report.pdf" target="_blank">📔 Technical Report</a>
   &nbsp;|
   <br/>
@@ -46,12 +47,12 @@ In this work, we present MiMo-7B, a series of models trained from scratch and bo
 </p>
 We open-source MiMo-7B series, including checkpoints of the base model, SFT model, RL model trained from base model, and RL model trained from the SFT model.
-We believe this report along with the models will provides valuable insights to develop powerful reasoning LLM that benefit the larger community.
 ### 🌟 Highlights
 - **Pre-Training: Base Model Born for Reasoning**
-  - We optimize data preprocessing pipeline, enhancing text extraction toolkits and applying multi-dimensional data filtering to increase reasoning pattern density in pre-training data. We also employ multiple strategies to generate massive diverse synthetic reasoning data.
   - We adopt a three-stage data mixture strategy for pre-training. Overall, MiMo-7B-Base is pre-trained on approximately 25 trillion tokens.
   - We incorporate Multiple-Token Prediction as an additional training objective, which enhances model performance and accelerates inference.
@@ -60,62 +61,83 @@ We believe this report along with the models will provides valuable insights to
     - To mitigate the sparse reward issue for challenging code problems, we introduce a test difficulty driven code reward. By assigning fine-grained scores for test cases with varying difficulty levels, the policy can be more effectively optimized via dense reward signal.
     - We implement a data re-sampling strategy for easy problems to enhance rollout sampling efficiency and stabilize policy updates, particularly in the later phases of RL training.
-- **RL Infrastructures**
-    - We develop a Seamless Rollout Engine to accelerate RL training and validation. Our design integrates continuous rollout, asynchronous reward computation, and early termination to minimize GPU idle time, achieving 2.29 \\(\times\\) faster training and 1.96 \\(\times\\) faster validation.
-    - We support MTP in vLLM and enhance the robustness of the inference engine in RL system.
 ## II. Model Details
-> Models are avaliable at [https://huggingface.co/XiaomiMiMo](https://huggingface.co/XiaomiMiMo)
-|    **Model**    |                                **Description**                                |                                   **Download**                                    |
-| :-------------: | :---------------------------------------------------------------------------: | :-------------------------------------------------------------------------------: |
-|  MiMo-7B-Base   |               Base model with extraordinary reasoning potential               |    [🤗 XiaomiMiMo/MiMo-7B-Base](https://huggingface.co/XiaomiMiMo/MiMo-7B-Base)    |
-| MiMo-7B-RL-Zero |                       RL model trained from base model                        | [🤗 XiaomiMiMo/MiMo-7B-RL-Zero](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-Zero) |
-|   MiMo-7B-SFT   |                       SFT model trained from base model                       |     [🤗 XiaomiMiMo/MiMo-7B-SFT](https://huggingface.co/XiaomiMiMo/MiMo-7B-SFT)     |
-| **MiMo-7B-RL**  | RL model trained from SFT model, superior performance matching OpenAI o1-mini |      [🤗 XiaomiMiMo/MiMo-7B-RL](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL)      |
 ## III. Evaluation Results
-| Benchmark                     | GPT-4o-0513 | Claude-3.5-Sonnet-1022 | OpenAI o1-mini | QwQ-32B-Preview | R1-Distill-Qwen-14B | R1-Distill-Qwen-7B | **MiMo-7B-RL** |
-| ----------------------------- | :---------: | :--------------------: | :------------: | :-------------: | :-----------------: | :----------------: | :------------: |
-| **General**                   |             |                        |                |                 |                     |                    |                |
-| GPQA Diamond<br/>(Pass@1)     |    49.9     |          65.0          |      60.0      |      54.5       |        59.1         |        49.1        |      54.4      |
-| SuperGPQA<br/>(Pass@1)        |    42.4     |          48.2          |      45.2      |      43.6       |        40.6         |        28.9        |      40.5      |
-| DROP<br/>(3-shot F1)          |    83.7     |          88.3          |      83.9      |      71.2       |        85.5         |        77.0        |      78.7      |
-| MMLU-Pro<br/>(EM)             |    72.6     |          78.0          |      80.3      |      52.0       |        68.8         |        53.5        |      58.6      |
-| IF-Eval<br/>(Prompt Strict)   |    84.3     |          86.5          |      84.8      |      40.4       |        78.3         |        60.5        |      61.0      |
-| **Mathematics**               |             |                        |                |                 |                     |                    |                |
-| MATH-500<br/>(Pass@1)         |    74.6     |          78.3          |      90.0      |      90.6       |        93.9         |        92.8        |      95.8      |
-| AIME 2024<br/>(Pass@1)        |     9.3     |          16.0          |      63.6      |      50.0       |        69.7         |        55.5        |      68.2      |
-| AIME 2025<br/>(Pass@1)        |    11.6     |          7.4           |      50.7      |      32.4       |        48.2         |        38.8        |      55.4      |
-| **Code**                      |             |                        |                |                 |                     |                    |                |
-| LiveCodeBench v5<br/>(Pass@1) |    32.9     |          38.9          |      53.8      |      41.9       |        53.1         |        37.6        |      57.8      |
-| LiveCodeBench v6<br/>(Pass@1) |    30.9     |          37.2          |      46.8      |      39.1       |        31.9         |        23.9        |      49.3      |
 MiMo-7B series
-| Benchmark                     | MiMo-7B-Base | MiMo-7B-RL-Zero | MiMo-7B-SFT | **MiMo-7B-RL** |
-| ----------------------------- | :----------: | :-------------: | :---------: | :------------: |
-| **Mathematics**               |              |                 |             |                |
-| MATH500<br/>(Pass@1)          |     37.4     |      93.6       |    93.0     |      95.8      |
-| AIME 2024<br/>(Pass@1)        |     32.9     |      56.4       |    58.7     |      68.2      |
-| AIME 2025<br/>(Pass@1)        |     24.3     |      46.3       |    44.3     |      55.4      |
-| **Code**                      |              |                 |             |                |
-| LiveCodeBench v5<br/>(Pass@1) |     32.9     |      49.1       |    52.3     |      57.8      |
-| LiveCodeBench v6<br/>(Pass@1) |     29.1     |      42.9       |    45.5     |      49.3      |
 > [!IMPORTANT]
-> The evaluation are conducted with `temperature=0.6`.
 >
 > AIME24 and AIME25 are with averaged score of 32 repetitions. LiveCodeBench v5 (20240801-20250201), LiveCodeBench v6 (20250201-20250501), GPQA-Diamond and IF-Eval are with averaged score of 8 repetitions. MATH500 and SuperGPQA are with a single run.
 ## IV. Deployment
 ### vLLM inference
-1. [Recommended] We official support inference with MiMo-MTP using [our fork of vLLM](https://github.com/XiaomiMiMo/vllm/tree/feat_mimo_mtp).
 Example script
@@ -180,9 +202,9 @@ Example script
 ```py
 from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer
-model_path = "/path/to/MiMo"
-model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
-tokenizer = AutoTokenizer.from_pretrained(model_path)
 inputs = tokenizer(["Today is"], return_tensors='pt')
 output = model.generate(**inputs, max_new_tokens = 100)
 print(tokenizer.decode(output.tolist()[0]))
@@ -190,7 +212,7 @@ print(tokenizer.decode(output.tolist()[0]))
 ### Recommended environment and prompts
-- We recommend using [our fork of vLLM](https://github.com/XiaomiMiMo/vllm/tree/feat_mimo_mtp) which is developed based on vLLM 0.7.3.
 - We recommend using empty system prompt.
 > We haven't verified MiMo with other inference engines and welcome contributions based on the model definition in the Huggingface repo 💻.
@@ -210,4 +232,4 @@ print(tokenizer.decode(output.tolist()[0]))
 ## VI. Contact
-Please contact us at [mimo@xiaomi.com](mailto:mimo@xiaomi.com) or open an issue if you have any questions.

 ---
 license: mit
 ---
 <div align="center">
   <picture>
     <source srcset="https://github.com/XiaomiMiMo/MiMo/raw/main/figures/Xiaomi_MiMo_darkmode.png?raw=true" media="(prefers-color-scheme: dark)">
 <h3 align="center">
   <b>
+    <span>━━━━━━━━━━━━━━━━━━━━━━━━━</span>
     <br/>
     Unlocking the Reasoning Potential of Language Model<br/>From Pretraining to Posttraining
     <br/>
+    <span>━━━━━━━━━━━━━━━━━━━━━━━━━</span>
     <br/>
   </b>
 </h3>
   |
   <a href="https://huggingface.co/XiaomiMiMo" target="_blank">🤗 HuggingFace</a>
   &nbsp;|
+  <a href="https://www.modelscope.cn/organization/XiaomiMiMo" target="_blank">🤖️ ModelScope</a>
+  &nbsp;|
   <a href="https://github.com/XiaomiMiMo/MiMo/blob/main/MiMo-7B-Technical-Report.pdf" target="_blank">📔 Technical Report</a>
   &nbsp;|
   <br/>
 </p>
 We open-source MiMo-7B series, including checkpoints of the base model, SFT model, RL model trained from base model, and RL model trained from the SFT model.
+We believe this report along with the models will provide valuable insights to develop powerful reasoning LLMs that benefit the larger community.
 ### 🌟 Highlights
 - **Pre-Training: Base Model Born for Reasoning**
+  - We optimize the data preprocessing pipeline, enhancing text extraction toolkits and applying multi-dimensional data filtering to increase reasoning pattern density in pre-training data. We also employ multiple strategies to generate massive diverse synthetic reasoning data.
   - We adopt a three-stage data mixture strategy for pre-training. Overall, MiMo-7B-Base is pre-trained on approximately 25 trillion tokens.
   - We incorporate Multiple-Token Prediction as an additional training objective, which enhances model performance and accelerates inference.
     - To mitigate the sparse reward issue for challenging code problems, we introduce a test difficulty driven code reward. By assigning fine-grained scores for test cases with varying difficulty levels, the policy can be more effectively optimized via dense reward signal.
     - We implement a data re-sampling strategy for easy problems to enhance rollout sampling efficiency and stabilize policy updates, particularly in the later phases of RL training.
+- **RL Infrastructure**
+    - We develop a Seamless Rollout Engine to accelerate RL training and validation. Our design integrates continuous rollout, asynchronous reward computation, and early termination to minimize GPU idle time, achieving $2.29\times$ faster training and $1.96\times$ faster validation.
+    - We support MTP in vLLM and enhance the robustness of the inference engine in the RL system.
 ## II. Model Details
+The MTP layers of MiMo-7B is tuned during pretraining and SFT and freezed during RL. With one MTP layer for speculative decoding, the acceptance rate is about 90%.
+<p align="center">
+  <img width="80%" src="https://github.com/XiaomiMiMo/MiMo/raw/main/figures/architecture.png?raw=true">
+</p>
+> Models are available at [https://huggingface.co/XiaomiMiMo](https://huggingface.co/XiaomiMiMo) and [https://www.modelscope.cn/organization/XiaomiMiMo](https://www.modelscope.cn/organization/XiaomiMiMo)
+|    **Model**    |                                **Description**                                |                            **Download (HuggingFace)**                             |                                  **Download (ModelScope)**                                  |
+| :-------------: | :---------------------------------------------------------------------------: | :-------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------: |
+|  MiMo-7B-Base   |               Base model with extraordinary reasoning potential               |    [🤗 XiaomiMiMo/MiMo-7B-Base](https://huggingface.co/XiaomiMiMo/MiMo-7B-Base)    |    [🤖️ XiaomiMiMo/MiMo-7B-Base](https://www.modelscope.cn/models/XiaomiMiMo/MiMo-7B-Base)    |
+| MiMo-7B-RL-Zero |                       RL model trained from base model                        | [🤗 XiaomiMiMo/MiMo-7B-RL-Zero](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL-Zero) | [🤖️ XiaomiMiMo/MiMo-7B-RL-Zero](https://www.modelscope.cn/models/XiaomiMiMo/MiMo-7B-RL-Zero) |
+|   MiMo-7B-SFT   |                       SFT model trained from base model                       |     [🤗 XiaomiMiMo/MiMo-7B-SFT](https://huggingface.co/XiaomiMiMo/MiMo-7B-SFT)     |     [🤖️ XiaomiMiMo/MiMo-7B-SFT](https://www.modelscope.cn/models/XiaomiMiMo/MiMo-7B-SFT)     |
+|   MiMo-7B-RL    | RL model trained from SFT model, superior performance matching OpenAI o1-mini |      [🤗 XiaomiMiMo/MiMo-7B-RL](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL)      |      [🤖️ XiaomiMiMo/MiMo-7B-RL](https://www.modelscope.cn/models/XiaomiMiMo/MiMo-7B-RL)      |
 ## III. Evaluation Results
+| Benchmark                     | GPT-4o-0513 | Claude-3.5-Sonnet-1022 | OpenAI o1-mini | QwQ-32B-Preview | R1-Distill-Qwen-14B | R1-Distill-Qwen-7B | MiMo-7B-RL |
+| ----------------------------- | :---------: | :--------------------: | :------------: | :-------------: | :-----------------: | :----------------: | :--------: |
+| **General**                   |             |                        |                |                 |                     |                    |            |
+| GPQA Diamond<br/>(Pass@1)     |    49.9     |          65.0          |      60.0      |      54.5       |        59.1         |        49.1        |    54.4    |
+| SuperGPQA<br/>(Pass@1)        |    42.4     |          48.2          |      45.2      |      43.6       |        40.6         |        28.9        |    40.5    |
+| DROP<br/>(3-shot F1)          |    83.7     |          88.3          |      83.9      |      71.2       |        85.5         |        77.0        |    78.7    |
+| MMLU-Pro<br/>(EM)             |    72.6     |          78.0          |      80.3      |      52.0       |        68.8         |        53.5        |    58.6    |
+| IF-Eval<br/>(Prompt Strict)   |    84.3     |          86.5          |      84.8      |      40.4       |        78.3         |        60.5        |    61.0    |
+| **Mathematics**               |             |                        |                |                 |                     |                    |            |
+| MATH-500<br/>(Pass@1)         |    74.6     |          78.3          |      90.0      |      90.6       |        93.9         |        92.8        |    95.8    |
+| AIME 2024<br/>(Pass@1)        |     9.3     |          16.0          |      63.6      |      50.0       |        69.7         |        55.5        |    68.2    |
+| AIME 2025<br/>(Pass@1)        |    11.6     |          7.4           |      50.7      |      32.4       |        48.2         |        38.8        |    55.4    |
+| **Code**                      |             |                        |                |                 |                     |                    |            |
+| LiveCodeBench v5<br/>(Pass@1) |    32.9     |          38.9          |      53.8      |      41.9       |        53.1         |        37.6        |    57.8    |
+| LiveCodeBench v6<br/>(Pass@1) |    30.9     |          37.2          |      46.8      |      39.1       |        31.9         |        23.9        |    49.3    |
 MiMo-7B series
+| Benchmark                     | MiMo-7B-Base | MiMo-7B-RL-Zero | MiMo-7B-SFT | MiMo-7B-RL |
+| ----------------------------- | :----------: | :-------------: | :---------: | :--------: |
+| **Mathematics**               |              |                 |             |            |
+| MATH500<br/>(Pass@1)          |     37.4     |      93.6       |    93.0     |    95.8    |
+| AIME 2024<br/>(Pass@1)        |     32.9     |      56.4       |    58.7     |    68.2    |
+| AIME 2025<br/>(Pass@1)        |     24.3     |      46.3       |    44.3     |    55.4    |
+| **Code**                      |              |                 |             |            |
+| LiveCodeBench v5<br/>(Pass@1) |     32.9     |      49.1       |    52.3     |    57.8    |
+| LiveCodeBench v6<br/>(Pass@1) |     29.1     |      42.9       |    45.5     |    49.3    |
 > [!IMPORTANT]
+> The evaluations are conducted with `temperature=0.6`.
 >
 > AIME24 and AIME25 are with averaged score of 32 repetitions. LiveCodeBench v5 (20240801-20250201), LiveCodeBench v6 (20250201-20250501), GPQA-Diamond and IF-Eval are with averaged score of 8 repetitions. MATH500 and SuperGPQA are with a single run.
 ## IV. Deployment
+### SGLang Inference
+Thanks to the [contribution](https://github.com/sgl-project/sglang/pull/5921) from the SGLang team, we supported MiMo in SGLang mainstream within 24h with MTP coming soon.
+Example Script
+```bash
+# Install the latest SGlang from main branch
+python3 -m uv pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git/@main#egg=sglang&subdirectory=python"
+# Launch SGLang Server
+python3 -m sglang.launch_server --model-path XiaomiMiMo/MiMo-7B-RL --host 0.0.0.0 --trust-remote-code
+```
+Detailed usage can be found in [SGLang documents](https://docs.sglang.ai/backend/send_request.html). MTP will also be supported in 24h.
 ### vLLM inference
+1. [Recommended] We officially support inference with MiMo-MTP using [our fork of vLLM](https://github.com/XiaomiMiMo/vllm/tree/feat_mimo_mtp_stable_073).
 Example script
 ```py
 from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer
+model_id = "XiaomiMiMo/MiMo-7B-Base"
+model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(model_id)
 inputs = tokenizer(["Today is"], return_tensors='pt')
 output = model.generate(**inputs, max_new_tokens = 100)
 print(tokenizer.decode(output.tolist()[0]))
 ### Recommended environment and prompts
+- We recommend using [our fork of vLLM](https://github.com/XiaomiMiMo/vllm/tree/feat_mimo_mtp_stable_073) which is developed based on vLLM 0.7.3.
 - We recommend using empty system prompt.
 > We haven't verified MiMo with other inference engines and welcome contributions based on the model definition in the Huggingface repo 💻.
 ## VI. Contact
+Please contact us at [mimo@xiaomi.com](mailto:mimo@xiaomi.com) or open an issue if you have any questions.