Bee - A full-stack, multimodal, large-scale model solution jointly open-sourced by Tsinghua University and Tencent.

What is Bee?

Bee is a high-quality multimodal large language model (MLLM) project jointly launched by Tsinghua University and Tencent's Hunyuan team, addressing the performance bottleneck caused by insufficient data quality in open-source models. The project's core contributions include: Honey-Data-15M, a high-quality supervised fine-tuning dataset containing approximately 15 million question-answer pairs, which improves data quality through multi-step cleaning and two-layer thought chain (CoT) augmentation strategies; HoneyPipe and DataStudio, open-source data preparation pipelines and frameworks providing transparent and reproducible data processing methods; and the Bee-8B model, an 8B-parameter model trained on Honey-Data-15M, which has broken the state-of-the-art (SOTA) records for fully open-source MLLMs in multiple benchmark tests, achieving performance comparable to or even better than some semi-open-source models.

Bee's main functions

• Building high-quality datasetsHoney-Data-15M has been released, a supervised fine-tuning dataset of 15 million data points that has been meticulously cleaned and augmented with a two-layer thought chain (CoT), significantly improving data quality and providing a solid foundation for training large multimodal models.
• Full-stack data processing pipelineHoneyPipe and DataStudio are open-source tools that provide a complete data processing approach, from data aggregation and noise filtering to CoT enhancement, ensuring transparency and reproducibility in data preparation, surpassing traditional static dataset publishing models.
• High-performance model training and validationThe Bee-8B model trained on Honey-Data-15M broke the performance record of fully open-source multimodal large language models in multiple benchmark tests, demonstrating the key role of high-quality data in improving model capabilities.
• Building an open source ecosystemIt provides complete open-source resources, including datasets, data processing pipelines, training recipes, evaluation tools, and model weights, to promote the development of the open-source community and assist academia and developers in research and application in the field of multimodal large models.

Bee's technical principles

• Data aggregation and deduplicationCollect a large number of image-text pairs from multiple data sources and ensure data diversity and efficient processing through strict deduplication.
• Noise filteringBy combining rules and models, we can remove noisy data with incorrect formatting, low-quality images, or mismatched instructions, thereby improving data quality.
• Mind Chain (CoT) ExpansionBy employing both short CoT and long CoT strategies, detailed reasoning processes are generated for instructions of varying complexity, thereby enhancing the model's reasoning capabilities.
• Fidelity verificationSemantic comparisons are performed using a validation model (LLM-as-a-Judge) to ensure the correctness and consistency of the generated CoT responses.
• Model training and optimizationThe Bee-8B model was trained on the high-quality dataset Honey-Data-15M, and its performance was optimized through techniques such as supervised fine-tuning (SFT) and reinforcement learning (RL).

Bee's project address

• Project official websitehttps://open-bee.github.io/
• HuggingFace model libraryhttps://huggingface.co/collections/Open-Bee/bee
• arXiv technical paperhttps://arxiv.org/pdf/2510.13795
• Honey-Data-15M datasethttps://huggingface.co/datasets/Open-Bee/Honey-Data-15M

Bee's application scenarios

• Multimodal content generationIt is used to generate high-quality image descriptions, video captions, etc., improving the efficiency and diversity of content creation.
• Intelligent question answering systemIn solving complex problems, it leverages its powerful reasoning capabilities to provide users with accurate and detailed answers.
• EducationIt assists in teaching, generates teaching materials or answers student questions, and supports personalized learning.
• Research supportIt helps researchers organize and analyze data to generate research reports or experimental design suggestions.
• Business IntelligenceAnalyze market trends and user feedback to provide data support and forecasts for decision-making.
• HealthcareIt assists in medical diagnosis, generates medical image analysis reports, or provides medical consultation advice.