DeepSeek-V2-Lite-BitsMoE-2bit

BitsMoE-DeepSeek-V2-Lite-2bit

This repository provides a 2-bit quantized version of DeepSeek-V2-Lite using the BitsMoE quantization framework. The model was introduced in the paper BitsMoE: Cost-Aware Bit Allocation in Spectral Space for MoE LLM Quantization.

The model is based on deepseek-ai/DeepSeek-V2-Lite and is intended for efficient inference of Mixture-of-Experts (MoE) large language models with significantly reduced memory footprint.

• Paper: BitsMoE: Cost-Aware Bit Allocation in Spectral Space for MoE LLM Quantization
• Code: GitHub - zjiayu064/BitsMoE

Model Details

• Base model: deepseek-ai/DeepSeek-V2-Lite
• Quantization: 2-bit (Experts-Only)
• Architecture: Mixture-of-Experts (MoE)
• Method: BitsMoE, a cost-aware bit-allocation framework for MoE LLM quantization.

Usage

Since this model uses custom architecture code via auto_map, ensure you load it with trust_remote_code=True.

Please refer to the BitsMoE GitHub repository for environment setup, custom kernels, and detailed usage instructions.

By default, models are downloaded from Hugging Face via the BitsMoE CLI:

bitsmoe demo

To run evaluation on benchmarks like MMLU or Hellaswag using the provided configs:

bitsmoe eval --config configs/deepseekv2/eval.yaml

License

This model follows the license of the original DeepSeek-V2-Lite model. Please refer to the license information provided in the metadata and the original DeepSeek model license for details.

Citation

If you find our work useful, please consider citing:

@misc{zhao2026bitsmoecostawarebitallocation,
      title={BitsMoE: Cost-Aware Bit Allocation in Spectral Space for MoE LLM Quantization}, 
      author={Jiayu Zhao and Zihan Teng and Minhao Fan and Tianrui Ma and Wentao Ren and Song Chen and Weichen Liu},
      year={2026},
      eprint={2606.00079},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2606.00079}, 
}