OpenBMBTranslation site

In today’s AI field, large models have become a crucial engine driving technological progress. However, training, fine-tuning, and inference of models with over 10 billion parameters pose significant demands on computational resources and technical capabilities, deterring many developers. To address this challenge, Tsinghua University’s Natural Language Processing Laboratory and the Beijing Academy of Artificial Intelligence‘s Large Model Acceleration Technology Innovation Center jointly initiated the OpenBMB open-source community, aiming to build a large-scale pre-trained language model library and related tools, lowering the threshold for large model usage and accelerating their application.

Website Introduction

OpenBMB focuses on providing end-to-end tools for large model training, fine-tuning, and inference, targeting global developers, especially individuals and teams interested in large models but limited by resources.

Key Features

• BMTrain: Efficient large model training tool, saving up to 90% in training costs compared to frameworks like DeepSpeed.
• BMCook: Large model compression toolkit, enhancing inference speed by 10 times through techniques like quantization, pruning, and distillation, while maintaining over 90% of the original performance.
• BMInf: Low-cost efficient inference tool, enabling inference of large models with over 10 billion parameters using affordable GPUs.
• OpenPrompt: Provides a unified interface for prompt learning template language, facilitating quick deployment of prompt learning methods.

Related Projects

The OpenBMB team has also released tools like OpenKE, OpenNRE, and OpenNE, collectively garnering over 58,000 stars on GitHub, ranking 148th among global institutions.

Advantages

OpenBMB’s toolchain significantly lowers the threshold for using large models, allowing developers with limited resources to engage in large model research and applications. Its efficient training and inference tools help users reduce costs while enhancing model performance.

Pricing

All tools provided by OpenBMB are open-source under the Apache License 2.0 and are free for developers to use.

Summary

OpenBMB was jointly initiated by Tsinghua University and the Beijing Academy of Artificial Intelligence in 2022, located in China, dedicated to providing a large-scale pre-trained language model library and related tools. Through these innovative features, users can train, fine-tune, and infer large models more cost-effectively and efficiently, accelerating the deployment of AI applications.