jedisct1/MiMo-7B-RL-GGUF

MiMo-7B-RL (GGUF)

This is a GGUF quantized version of XiaomiMiMo/MiMo-7B-RL, optimized for use with llama.cpp, Ollama, LM Studio, and other GGUF-compatible inference engines. The model has been converted from the original SafeTensors format to GGUF.

Model Description

MiMo-7B-RL is a powerful 7B parameter language model developed by Xiaomi, specifically designed for enhanced reasoning capabilities in both mathematics and code. The original model matches the performance of OpenAI's o1-mini in many benchmarks.

Model Details

• Original Model: MiMo-7B-RL by Xiaomi
• Parameters: 7 billion
• Context Length: 32,768 tokens
• Architecture: Modified transformer with 36 layers, 32 attention heads
• Original Format: SafeTensors
• Converted Format: GGUF
• License: MIT

Key features of the original model:

• Trained using a specialized pre-training strategy focused on reasoning tasks
• Fine-tuned with reinforcement learning on 130K mathematics and code problems
• Demonstrates superior performance in both mathematical reasoning and coding tasks
• Matches performance of much larger models in reasoning capabilities

Usage

With Ollama

ollama run mimo-7b-rl-q8

With LM Studio

1. Load the model through the LM Studio interface
2. Select the GGUF file
3. Configure your desired settings
4. Start chatting!

With llama.cpp

./main -m mimo-7b-rl-q8.gguf -n 4096

Performance

The original model demonstrates impressive performance across various benchmarks:

Benchmark	Score
MATH-500 (Pass@1)	95.8%
AIME 2024 (Pass@1)	68.2%
AIME 2025 (Pass@1)	55.4%
LiveCodeBench v5 (Pass@1)	57.8%
LiveCodeBench v6 (Pass@1)	49.3%

Note: Performance metrics are from the original model. The GGUF conversion may show slightly different results due to quantization.

Limitations and Biases

The model inherits any limitations and biases present in the original MiMo-7B-RL model. Additionally:

• Q8 quantization may result in slightly reduced performance compared to the original model
• The model requires careful prompt engineering for optimal results in reasoning tasks
• Performance may vary depending on the specific GGUF inference implementation used

Training Details

The model was trained by Xiaomi using:

• Pre-training on approximately 25 trillion tokens
• Three-stage data mixture strategy
• Multiple-Token Prediction as an additional training objective
• RL fine-tuning on 130K mathematics and code problems

For detailed training information, please refer to the original model card.

Citation

If you use this model, please cite the original work:

@misc{xiaomi2025mimo,
      title={MiMo: Unlocking the Reasoning Potential of Language Model – From Pretraining to Posttraining},
      author={{Xiaomi LLM-Core Team}},
      year={2025},
      primaryClass={cs.CL},
      url={https://github.com/XiaomiMiMo/MiMo},
}

Acknowledgments

Original model development by Xiaomi LLM-Core Team.