SPES-9B

SPES-9B is a 9B-parameter Mixture-of-Experts (MoE) language model pretrained using SPES (SParse Expert Synchronization), a memory-efficient decentralized pretraining paradigm.

The model and framework were introduced in the paper: Pretraining A Large Language Model using Distributed GPUs: A Memory-Efficient Decentralized Paradigm.

Authors

Jinrui Zhang, Chaodong Xiao, Aoqi Wu, Xindong Zhang, and Lei Zhang.

Model Details

• Model name: SPES-9B
• Model type: Causal Mixture-of-Experts (MoE) language model
• Parameters: 9B (Upcycled from a dense Qwen3-1.7B checkpoint)
• Framework: SPES
• License: Apache-2.0

Description

SPES-9B was developed to address the memory constraints of GPU nodes in decentralized training environments. By training only a subset of experts per node, the SPES framework significantly reduces the memory footprint and eliminates the need for full-parameter transmission and high-speed cross-node interconnects. This allows the model to be trained effectively over standard internet connections while maintaining competitive performance compared to centralized baselines.

Project Links

• GitHub Repository: zjr2000/SPES
• Paper: Hugging Face Papers
• Training Logs: Weights & Biases

Intended Use

This model is intended for research on:

• Decentralized LLM pretraining paradigms.
• Mixture-of-Experts (MoE) training and synchronization mechanisms.
• Evaluation of pretrained language models trained under computational and bandwidth constraints.

Citation

If you use this model or the SPES framework in your research, please cite:

@article{zhang2026pretraining,
  title={Pretraining A Large Language Model using Distributed GPUs: A Memory-Efficient Decentralized Paradigm},
  author={Zhang, Jinrui and Xiao, Chaodong and Wu, Aoqi District and Zhang, Xindong and Zhang, Lei},
  journal={arXiv preprint arXiv:2602.11543},
  year={2026}
}

Acknowledgements

The SPES codebase is built upon the modeling and training infrastructure provided by OLMo (Allen Institute for AI) and utilizes MegaBlocks (Databricks) for efficient MoE operations.