Bee - Tencent hybrid open source full-stack multimodal large model project with Tsinghua University

What's Bee?

Bee is a full-stack open-source multimodal large model solution jointly launched by Tencent Mixed Element team and Tsinghua University, which narrows the performance gap between open-source and closed-source models by improving data quality. The project consists of three core achievements: Honey-Data-15M, a high-quality two-layer CoT dataset with a size of 15 million, the open-source data enhancement tools HoneyPipe and DataStudio, and Bee-8B, an 8B model trained based on the dataset, which outperforms mainstream semi-open-source models in multiple benchmarks, especially in mathematical reasoning and diagram comprehension tasks. outperforming mainstream semi-open source models. By making the dataset and methodology publicly available, the project provides the open source community with an important infrastructure for improving MLLM performance.

Features of Bee

• High-quality data sets: Provide Honey-Data-15M dataset expanded by fine cleaning and two-layer Chain of Thought (CoT), which significantly improves data quality and provides a solid foundation for model training.
• full-stack open source pipeline: Open source HoneyPipe and DataStudio, covering the entire process from data aggregation, noise filtering to inference enhancement, ensuring transparent and reproducible data processing.
• high performance model: The Bee-8B model, trained on high-quality data, has set a new performance record for all open-source multimodal large models in several benchmark tests, demonstrating excellent reasoning and understanding capabilities.
• multimodal fusion: It supports fusion processing of multiple modalities such as image and text, and is suitable for multimodal application scenarios such as visual question and answer and image description generation.
• Reasoning Enhancement: Generating detailed reasoning processes for complex tasks through short CoT and long CoT strategies to enhance the performance of models in complex problem solving.
• community-driven: Build an open source ecosystem that provides datasets, tools, and model weights, encourages community participation and contributions, and promotes the continued development of the technology.
• Flexible deployment: Supports a variety of deployment methods, including local deployment and cloud deployment, to meet the needs of different users.
• Continuous optimization: Continuous model evolution and performance improvement through data contribution incentives and online learning paradigms.

Bee's core strengths

• Excellent data quality: A high-quality Honey-Data-15M dataset is constructed through multi-step cleaning and two-layer chain-of-thinking (CoT) expansion, which significantly improves the accuracy and inference depth of the data.
• full-stack open source transparency: Provides full-stack open source tools from data processing to model training, including HoneyPipe and DataStudio, ensuring transparency and reproducibility throughout the process.
• Model Performance Leadership: The Bee-8B model sets performance records for all open-source multimodal large models in several benchmark tests, demonstrating powerful reasoning and complex task processing capabilities.
• Outstanding reasoning skills: Short CoT and long CoT strategies are used to generate detailed reasoning processes for tasks of different complexity, significantly enhancing the logical reasoning capability of the model.
• Open Source Ecology Improvement: Provides complete open source resources including datasets, training recipes, evaluation tools and model weights to help academics and developers get up to speed quickly and further develop.

What is Bee's official website?

• Project website:: https://open-bee.github.io/
• HuggingFace Model Library:: https://huggingface.co/collections/Open-Bee/bee
• arXiv Technical Paper:: https://arxiv.org/pdf/2510.13795
• Honey-Data-15M dataset:: https://huggingface.co/datasets/Open-Bee/Honey-Data-15M

Who Bee is for

• Artificial intelligence researchers: High-quality datasets and open-source models can be utilized for research and innovation in multimodal macromodeling.
• Developers and engineers: The ability to leverage open-source tools and models for application development and rapid integration of multimodal functionality.
• data scientist: Data can be processed and analyzed with HoneyPipe and DataStudio to improve data quality and model performance.
• educator: The Bee model can be utilized to generate instructional materials or assist in teaching and learning to enhance teaching and learning.
• content creator: You can quickly create high-quality graphic and video content with the help of multimodal content generation.
• business user: Bee models can be applied to intelligent customer service, market analysis, business intelligence and other scenarios to improve business efficiency.