Mimo 7B RL Zero

🚀 MiMo-7B: Unlocking the Reasoning Potential of Language Model

This project presents the MiMo-7B series of models, which are trained from scratch and designed for reasoning tasks. These models show extraordinary reasoning potential and superior performance in mathematics and code reasoning tasks.

🚀 Quick Start

This README provides an overview of the MiMo-7B series of models, including their pre - training and post - training strategies, evaluation results, and deployment methods. You can access the models on HuggingFace and ModelScope, and follow the deployment instructions to use them.

✨ Features

Pre - Training: Base Model Born for Reasoning

• Optimize the data preprocessing pipeline, enhance text extraction toolkits, and apply multi - dimensional data filtering to increase reasoning pattern density in pre - training data. Generate massive diverse synthetic reasoning data using multiple strategies.
• Adopt a three - stage data mixture strategy for pre - training. The MiMo - 7B - Base is pre - trained on approximately 25 trillion tokens.
• Incorporate Multiple - Token Prediction as an additional training objective to enhance model performance and accelerate inference.

Post - Training Recipe: Pioneering Reasoning Model

• Curate 130K mathematics and code problems as RL training data, verified by rule - based verifiers. Each problem is carefully cleaned and its difficulty is assessed. Only rule - based accuracy rewards are used to avoid potential reward hacking.
• Introduce a test difficulty driven code reward to mitigate the sparse reward issue for challenging code problems. Assign fine - grained scores for test cases with varying difficulty levels to optimize the policy effectively.
• Implement a data re - sampling strategy for easy problems to enhance rollout sampling efficiency and stabilize policy updates, especially in the later phases of RL training.

RL Infrastructure

• Develop a Seamless Rollout Engine to accelerate RL training and validation. Integrate 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.
• Support MTP in vLLM and enhance the robustness of the inference engine in the RL system.

📦 Installation

SGLang Inference

# 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-Zero --host 0.0.0.0 --trust-remote-code

vLLM inference

1. [Recommended] Install our fork of vLLM:

# No specific installation command provided in the original, assume cloning the repo
git clone https://github.com/XiaomiMiMo/vllm/tree/feat_mimo_mtp_stable_073

2. For registering a vLLM loader without MTP parameters:

# Assume cloning the repo to get the registration script
git clone https://github.com/XiaomiMiMo/MiMo.git

HuggingFace inference

pip install transformers

💻 Usage Examples

SGLang Inference

# 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-Zero --host 0.0.0.0 --trust-remote-code

Detailed usage can be found in SGLang documents. MTP will also be supported in 24h.

vLLM inference

Basic Usage

from vllm import LLM, SamplingParams

model_path = "/path/to/MiMo"
llm = LLM(
    model=model_path,
    trust_remote_code=True,
    num_speculative_tokens=1,
    disable_log_stats=False
)
sampling_params = SamplingParams(temperature=0.6)

conversation = [
    {
        "role": "system",
        "content": ""
    },
    {
        "role": "user",
        "content": "Write an essay about the importance of higher education.",
    },
]

outputs = llm.chat(conversation,
                   sampling_params=sampling_params,
                   use_tqdm=False)

for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

print("=" * 80)

Advanced Usage

import register_mimo_in_vllm

from vllm import LLM, SamplingParams

model_path = "/path/to/MiMo"
llm = LLM(
    model=model_path,
    trust_remote_code=True,
    # num_speculative_tokens=1,
    disable_log_stats=False
)
sampling_params = SamplingParams(temperature=0.6)

HuggingFace inference

from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer

model_id = "XiaomiMiMo/MiMo-7B-RL-Zero"
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(*

📚 Documentation

Model Details

The MTP layers of MiMo - 7B are tuned during pretraining and SFT and frozen during RL. With one MTP layer for speculative decoding, the acceptance rate is about 90%.

Evaluation Results

MiMo - 7B series

⚠️ Important Note

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.

📄 License

This model repository is licensed under the MIT License.